Last updated on March 10th, 2026 at 04:17 am
Update #12 /loop (agentic,but carefully limited), code review and a lawsuit, with friends : Anthropic presses on
The March 7 release of Claude Code’s /loop command and cron scheduling represents something genuinely interesting from an agentic design perspective: deliberate limits as a feature. Users can set up recurring background tasks (checking error logs, monitoring pull requests, generating morning Slack summaries) at fixed intervals using standard cron expressions, with up to 50 scheduled tasks per session. But the tasks auto-delete after three days. This is intentional design philosophy meeting degradation limits, running counter to the “set it and forget it” ambition driving most agentic AI development.
The three-day expiry makes /loop a case study in what we could call atomized agency: limited, time-restricted, self-terminating task execution that resists the drift toward persistent autonomous systems. Where most of the industry is racing to build agents that run indefinitely across long horizons, Anthropic has shipped something closer to a controlled burn: useful, scoped, and inherently temporary. This fits neatly into the broader pattern of Anthropic’s agentic rollout (agent teams, Cowork, computer use) where the throughline is always capability with guardrails rather than capability at all costs. For enterprise buyers evaluating agentic reliability, the distinction matters enormously.
Code Review, which launched two days later on March 9, extends this philosophy in a different direction. Rather than constraining time horizons, it limits authority. When a PR opens, Code Review dispatches multiple specialized agents in parallel; some hunting for bugs, others verifying findings to suppress false positives, a final pass ranking issues by severity. The system won’t approve PRs; that decision stays with humans. What it does is close the gap between what’s shipping and what’s actually being read. Internally at Anthropic, the share of PRs receiving substantive review comments jumped from 16% to 54% after deployment. On large PRs exceeding 1,000 lines, 84% produce findings averaging 7.5 issues; on small PRs under 50 lines, that drops to 31%. The design choice to focus exclusively on logic errors rather than style nitpicks is telling. Anthropic deliberately scoped the tool to surface only high-priority, immediately actionable findings. At $15–25 per review billed on token usage, it’s priced as a serious enterprise tool, not a developer toy. This is agent-team architecture pointed at a real bottleneck: the vibe-coding explosion has made writing code cheap and fast, but review capacity hasn’t scaled to match. Code Review doesn’t solve that by removing humans from the loop, it solves it by making the human reviewer’s remaining attention count for more. Combined with /loop’s temporal constraints, the pattern is consistent: Anthropic keeps shipping agentic capability that exemplifies their design philosophy: pushing capabilities forward, it carefully, hitting the sweet spot between real productivity and limited autonomy, deliberately refusing to be the final authority.
The company also filed lawsuits this week challenging a U.S. Defense Department decision to designate the company a “supply-chain risk,” with amicus briefs from dozens of researchers from OpenAI and Google DeepMind, submitted supporting Anthropic, warning that the government’s approach could chill AI innovation and set a precedent allowing federal agencies to pressure companies into altering safety policies. The intervention by employees from rival labs, reportedly including prominent figures such as Google’s chief scientist Jeff Dean, signals that the company is prepared to go the distance on principle, and has substantial industry support in doing so.
The Claude app remains in first position in the App Store.
UPDATE #11: They’re Number One, plus new tools, one openly competitive.
Anthropic pushes on, gaining symbolic and real momentum. In the wake of CEO Amodei’s refusal to back down on military use principle, the app hit number one in the AppStore for the first time, while OpenAI contents with a backlash and a boycott campaign, which CEO Sam Altman’s AMA (ask me anything) session on Twitter/X seems to have done little to address, despite supportive comments from VC Paul Graham and others.
Last week the company launched Remote Control for Claude Code, and the implications go beyond convenience. The feature lets developers start a coding session in their desktop terminal and then pick it up from a phone, tablet, or browser, while everything continues running on the local machine. No code moves to the cloud. Your phone is just a window into the session still running on your computer, with full access to your filesystem and project configuration. It’s available now for Max subscribers ($100-$200/month) with Pro ($20/month) access coming soon.
The numbers behind it tell the real story: Claude Code has hit a $2.5 billion annualized run rate as of February 2026 (more than doubled since the start of the year) with 29 million daily installs in VS Code. An estimated 4% of all public GitHub commits worldwide are now authored by Claude Code. Remote Control doesn’t just untether developers from their desks. It signals that Anthropic is betting the local-first model will beat the cloud-first approach competitors like Cursor are pursuing — a meaningful architectural choice in an era when data sovereignty is becoming a strategic concern at every level.
Meanwhile, Anthropic launched what might be its most aggressive competitive move yet: a memory import tool at claude.com/import-memory that lets users transfer their stored memories from ChatGPT, Gemini, Copilot, or any other AI assistant directly into Claude. The process is almost absurdly simple: copy a prompt into your current AI, paste the output into Claude’s memory settings, done. No file exports, no API tokens, no JSON parsing. Imported memories include personal context, work details, technical preferences, communication style, and recurring tasks. It’s a direct play to eliminate the single biggest barrier to switching providers: the months of accumulated context that make starting over feel painful. OpenAI has a similar feature, a native feature that directly imports memories from other AI providers into ChatGPT in the way Anthropic’s import flow does. But the OpenAI export function is focused on letting a user download their own data (including memory entries) for their personal records, not a built-in cross-assistant memory migration tool and is not as explicitly competitive as Anthropic’s.
And then there’s the education play, which both companies are now running simultaneously. Anthropic launched its own Academy on Skilljar (anthropic.skilljar.com), offering 13 free courses with certificates covering everything from Claude 101 and the Claude API to Model Context Protocol, Claude Code, AI fluency foundations, and specialized tracks for educators and students. No Anthropic account required. No paywall. A post about it hit 3.2 million views on X in under 24 hours, with people calling it “probably the most valuable use of your time in 2026.” OpenAI is running the same play with OpenAI Academy: free courses, community events, hands-on learning directly inside ChatGPT, and formal certification programs piloted with Walmart, John Deere, Lowe’s, BCG, and Accenture. OpenAI’s stated goal: certify 10 million Americans by 2030.
Both companies have arrived at the same conclusion: the deepest moat in AI isn’t the model. It’s the workforce. Train millions of people on your tools, give them certificates they put on LinkedIn, and you’ve built switching costs that no technical benchmark can overcome. It’s not just about technology anymore, it’s about muscle memory.
UPDATE #10: The MCP Ecosystem Expands to the Free Tier and What It Actually Means
On the same day the Pentagon moved to strip safety guardrails from Claude, Anthropic expanded what its free tier can do in the opposite direction, opening access to pre-built connectors that let Claude plug directly into live tools like Google Drive, Google Calendar, Gmail, Notion, and WordPress. No paid plan required. Users authenticate through OAuth, and Claude can search, read, and interact with real data inside those systems. Paid tiers unlock the ability to add custom MCP server connections (which is where tools like the one below come in) but the base infrastructure is now available to everyone. Over 38 workplace integrations are live through Anthropic’s connector directory, with Salesforce, Snowflake, HubSpot, Jira, and others already connected. The MCP standard itself is heading toward a 1.0 stable release under Linux Foundation governance by mid-2026, and OpenAI is now building its own MCP-compatible connector system. In other words: AI models are no longer just generating text from training data. They are reaching into live business systems, pulling real information, and taking actions. Which makes the question of who controls these models, and what guardrails govern that access, not a theoretical debate but an operational one, as this week made violently clear.
Meanwgile, a post circulating widely this week claims that someone “gave Claude access to the entire stock market” through an open-source tool called Financial Datasets MCP Server. The tool is real. It uses the Model Context Protocol (an open standard Anthropic developed that lets external data sources plug directly into Claude) to pipe live financial statements, stock prices, crypto data, and company news into the model on demand. The data is structured, current, and sourced from an actual API, not generated from training data. That distinction matters: when Claude pulls an income statement through this pipeline, the numbers are real.
But the framing that this is a free Bloomberg terminal is marketing, not reality. Bloomberg terminals cost $50,000 a year because they provide real-time streaming, proprietary fixed-income pricing, derivatives modeling, a global messaging network, and analytics built over four decades. This tool gives you basic financials and historical prices through a free tier with rate limits. It is useful. It is not Bloomberg. More importantly, the reliability of the data pipeline does not solve the reliability problem at the reasoning layer. Claude can receive a perfectly accurate balance sheet and still miscalculate a debt-to-equity ratio, misidentify a trend, or produce a confident conclusion the numbers don’t support. The data is grounded. The analysis is still probabilistic.
That distinction is the kind of thing the exploitability gap obscures: the public sees “real data” and assumes “real answers,” when what they’re actually getting is real inputs processed by a system that still fails in the ways our research documents.
The broader MCP ecosystem is expanding fast. Alpha Vantage offers a free-tier financial data server. There are open-source MCP integrations for Google Drive, GitHub, Slack, databases, and more. Anthropic has built native integrations for Google Calendar and Gmail directly into claude.ai. This is the direction the industry is moving: AI models that don’t just generate text from training data but actively reach into live systems to pull real information. That is a genuine and significant shift.
It is also, in the context of everything happening this week, another reason why the question of who controls these models and what guardrails govern their access to live data systems is not abstract. It is the most concrete question in technology right now. It’s encouraging and somewhat heartening to see the democratization of technology, down to the free tier layer of the most sophisticated consumer technology product humanity has ever built, when so much of our world is moving in the other direction.
UPDATE #9: Amodei doesn’t cave under threats, but does retrench under competitive pressure
When Defense Secretary Pete Hegseth gave Dario Amodei a Friday deadline to strip all safeguards from Claude or face having Anthropic labeled a “supply chain risk” and potentially forced into compliance through the Defense Production Act, Amodei did something almost no one in tech does anymore: he held the line. In a public statement released Thursday, he wrote that the threats “do not change our position” and that Anthropic “cannot in good conscience” grant the Pentagon unrestricted access to its models. His two redlines — no fully autonomous weapons and no mass domestic surveillance of Americans — remained exactly where they were before the meeting, before the ultimatum, and before Under Secretary Emil Michael called him a “liar” with a “God complex” on X. Even former Trump AI adviser Dean Ball called the Pentagon’s dual threats “insane” and “incoherent.” Amodei’s response was measured, diplomatic, and immovable — the kind of pressure response that reveals what someone’s values are actually made of, not just what they say they are when the room is friendly.
That said, the picture isn’t entirely clean. The same week Amodei was staring down the Pentagon, Anthropic dropped the central pillar of its Responsible Scaling Policy: the 2023 commitment to halt model training if safety measures couldn’t be guaranteed in advance. The company framed the change as pragmatic: if Anthropic pauses training while OpenAI, xAI, and Google keep building, the companies with the weakest safety practices end up defining the frontier.
It’s a defensible position in a landscape where OpenAI dropped “safely” from its mission statement and xAI got approved for classified military use the same week it was generating nonconsensual deepfakes. But it does mean Anthropic is now operating with softer internal constraints at the exact moment its external stance is being tested hardest. What makes Amodei’s position noteworthy (and terrifying) is that he’s the only major AI CEO drawing any lines at all, and he’s holding those lines under threats that could genuinely damage his company’s trajectory.
One thing is clear: Amodei’s experience leaving OpenAI then being a distant, cautious second have prepared him for this moment. The changes he’s enacted at Anthropic are described by some as accidental, but being able to capitalize on emergent opportunities is a form of strategy in and of itself, and he’s demonstrated an ability to excel at that very clearly. They are now the most talked about AI company in the world, seeing the fastest growth in both revenue and productization, doing the most advanced work with the most advanced capabilities. They are literally rewriting the playbook.
UPDATE #8: Scheduled Tasks (with a caveat)
Cowork now supports scheduled tasks, letting Claude automatically handle recurring work at set times; things like daily morning briefs, weekly spreadsheet updates, or Friday team presentations.
The big caveat: Scheduled tasks only run while your computer is awake and the Claude Desktop app is open. If your computer is asleep or the app is closed when a task is scheduled to run, Cowork will skip it, then run it automatically once your computer wakes up or the app is reopened.
How to set it up: Tasks can be created from the “Scheduled” tab in the left sidebar, or by typing /schedule in the task command bar. Managing tasks — pausing, deleting, or running on demand — is done through the left sidebar. Plugins make it more powerful: Tasks can search Slack, query files, run web research, generate reports, and more — using any connectors and plugins you’ve set up in Cowork.
Availability: The feature is available on all paid plans (Pro, Max, Team, and Enterprise).
UPDATE #7: Agents and Claude for COBOL
IBM fell 14% on news that Anthropic highlighted (highlighted! already released! a PDF!) that its Claude Code tool can analyze and automate large parts of modernizing legacy COBOL systems, a programming language still widely used in banking, government and other mission-critical infrastructure. Investors reacted strongly, fearing this could reduce reliance on traditional consulting and legacy business tied to mainframes. As a result, IBM saw its worst daily decline in more than 25 years, amid broader tech and software sell-offs tied to AI disruption concerns, despite the fact that it was not news, nor would it affect IBM’s core business, it would even possibly strengthen its position in a growth market.
Anthropic also unveiled a major expansion of its enterprise AI agent strategy by launching a set of pre-built plugin-style agents designed to support specific business functions such as finance including analysis, investment banking, wealth management, engineering, legal, HR and design. These plugins are part of its broader Claude Cowork platform, enabling companies to deploy customizable Claude-powered agents within their own workflows. The new system aims to make it easy for corporate IT teams to roll out and tailor AI assistants with built-in skills for tasks like financial modeling, engineering specs, HR document creation, and more. Anthropic emphasizes that these aren’t one-size-fits-all bots but frameworks companies can modify to match their internal needs. The push positions Anthropic’s agent ecosystem as yet another a new layer in enterprise workflow tooling and orchestration and signals additional competition with traditional SaaS products in those domains.
On February 20, Anthropic also launched Claude Code Security, its first dedicated cybersecurity product. Unlike traditional static analysis tools that scan for known vulnerability patterns, Claude Code Security reviews entire codebases contextually, examining how components interact, how data flows through systems, and where architectural weaknesses create exploitable gaps. It’s code review the way a senior security engineer does it, not the way a linter does it.
The proof of concept was striking: using Opus 4.6, Anthropic’s internal team identified over 500 vulnerabilities in production open-source codebases — bugs that had survived years of expert human review. The company is now working through responsible disclosure with maintainers. CrowdStrike and Zscaler both dropped over 7% on the news, extending the pattern that has defined Anthropic’s 2026: enter a vertical, demonstrate capability, watch incumbent stocks move.
UPDATE #6: Claude Moves into macOS and Cross Platform, Cross-Function Dominance
No enterprise software company has done what Anthropic has done in the past sixty days: embedded a single AI model simultaneously across operating systems, legal platforms, design tools, communication infrastructure, financial data systems, and clinical environments, not as a native integration, but as the connective layer across other companies’ products.
Today, the company announced that it has expanded Claude Cowork to Windows and Intel-based macOS with full feature parity to the January release, meaning the same model, connectors, and skills system are now available across platforms. This is not just a UI port. It signals that Anthropic is standardizing the Cowork experience as a cross-platform desktop agent rather than a platform-specific experiment. For enterprise teams operating in mixed Windows–Mac environments, that is highly significant. It removes complexity from workflow, friction from deployment decisions and reinforces Cowork as a consistent layer for day-to-day operational work, regardless of OS.
Cowork operates within user-level permissions; writing and editing files, running commands, managing configurations, and interacting with connected systems when authorized. The shift is from chat-based assistance to local environment interaction: a desktop agent that can modify workflows, automate tasks, and reshape how users operate within their existing system constraints. That’s a meaningful architectural step toward agentic productivity.
For enterprise buyers, the cross-platform rollout reduces one of the most persistent barriers to AI deployment: environment fragmentation. Most organizations operate across Windows fleets, legacy Intel Macs, and increasingly Apple Silicon devices. A desktop agent that behaves consistently across those environments simplifies governance, rollout planning, training, and security review. More importantly, it positions Cowork not as a novelty interface but as an operational layer, one that can sit inside existing enterprise controls with document workflows, configuration management, and connected system tasks. The strategic shift is subtle but significant: AI is moving from advisory interface to embedded productivity infrastructure.
In practice, this means that when properly configured with enterprise-approved connectors and permissions, Claude Cowork can operate across multiple systems within the same workflow: drafting a contract stored locally, referencing Slack conversations, pulling structured data from a billing platform, updating documentation, or preparing materials for regulatory filings. The significance is coordinated system interaction. For enterprises, that orchestration layer — bridging operating systems, collaboration tools, design platforms, and structured business software, is where real, unprecedented productivity gains emerge.
It’s not plug and play; there’s integration work involved. But once that threshold is crossed, what’s on the other side is genuine single-pane synthesis: Claude not just orchestrating across systems, but seeing across them. That’s new.
UPDATE #5: 2/14: No caption necessary.
UPDATE #4: 2/10 Claude Cowork Moves into Windows and the Operating System
This is the post that will seemingly go on being updated indefinitely, tracking the evolution revolution. Today Anthropic announced integration of Claude into Windows. Not PowerPoint. Not Excel. Windows. The operating system. The original “innovation” (many say it borrowed heavily from Apple) that built the company, that runs everything. This is operational intelligence with access to everything including operating memory.
This expansion of Claude’s “Cowork” environment to Windows is less a platform update than a signal of architectural intent. With file access, multi-step task execution, plugins, and MCP connectors now available across operating systems, Anthropic is positioning Claude not as a conversational layer, but as an operational one. This moves the model decisively from advisory use into embedded execution, where it can read, write, and modify artifacts directly within a user’s working environment, rather than merely describing what should be done.
What matters most here is not autonomy in the abstract, but a. persistence and b. context. The introduction of global and folder-level instructions that carry across sessions effectively gives Claude a (scoped, policy-bound, but a …) memory. Its behavior can be shaped once and then reliably reproduced, with different limitations and roles depending on where it is operating. This turns the model into a repeatable actor within workflows, reducing the necessity of re-instruction and making it suitable for sustained, iterative work rather than isolated prompts.
In practice, this collapses a layer of coordination that has traditionally sat between intent and execution. Tasks that previously required translation (specs, handoffs, clarification cycles) can now be instantiated directly as actions and evaluated in place. This is not general autonomy, but limited agency: the model operates within explicit permissions, defined tools, and human-set objectives. Still.
The effect on productivity is real, particularly for the “glue work” that dominates knowledge labor but rarely justifies full engineering attention.
Seen in context, Cowork helps explain why roles across AI-first organizations are simultaneously widening in function but reducing in different function at the same time, getting compressed vs disappearing. When models can maintain context, execute sequences, and operate directly on real systems, the distinction between planning and doing becomes thinner. The risk, as always, lies not in the capability itself but in how it is integrated. Tools like this do not announce a sudden leap into the future; they normalize a new operating baseline. By the time the shift feels obvious, the coordination costs it removed are already gone.
What a time to be alive.
On February 5, Anthropic also released Claude Opus 4.6, its first major model launch of 2026. The headline feature is “agent teams”, multiple agents that can split a complex task into parallel workstreams and coordinate directly with each other, rather than one agent working sequentially. Opus 4.6 also expands Claude’s reach beyond its core developer audience into financial research, document production, and broader knowledge work. Anthropic’s head of product, Scott White, described the shift as moving from “vibe coding” to “vibe working.”
UPDATE #3: 2/4: The Legal Plug-ins
The surge continues. Anthropic has announced Legal Plugin for Claude Cowork (announced Feb 2-3, 2026 – yesterday/today), a move that caused a massive ripple effect on stock markets, (anticipating similar integration plays across hundreds of verticals). Anthropic’s specialized Legal Plugins as part of Claude Cowork (their agentic no-code tool) include:
Core Legal Capabilities:
∙ /review-contract – Clause-by-clause contract review against negotiation playbooks with green/yellow/red risk flags and redline suggestions
∙ /triage-nda – Categorizes NDAs for standard approval, counsel review, or full review
∙ /vendor-check – Check vendor agreement status
∙ /brief – Generate contextual briefings (daily briefs, topic research, incident response)
∙ /respond – Create templated responses for common inquiries (data subject requests, discovery holds)
Connects to: Slack, Box, Egnyte, Jira, Microsoft 365
Also Available: Midpage legal research integration (MCP connector) – this gives Claude access to comprehensive case law database and AI-powered citator
As one commentator said on Twitter/X: “RIP billable hour”
UPDATE #2: 2/4: The Work Tools Integration: Claude as Platform
Claude now integrates across a wide range of enterprise platforms spanning productivity suites, workflow management systems, vertical-specific data sources, and developer tools. These integrations are enabled through a standardized protocol rather than bespoke, one-off connections, allowing Claude to operate across multiple tools within a single task flow.
Claude’s Enterprise Integration Blitz
In a flex display of ecosystem expansion, Anthropic has transformed Claude from a standalone AI assistant into the connective tissue of enterprise productivity, rolling out integrations that span virtually every major business platform. These include:
Access and retrieval Systems Claude can read from, search, and summarize. These integrations expand visibility across documents, messages, datasets, and records, but do not grant execution authority.
Workflow coordination Systems where Claude can help sequence tasks, move information between tools, and assist with multi-step processes. This enables cross-system orchestration, but actions remain constrained, session-bound, and typically user-initiated.
Embedded assistance Environments where Claude operates directly inside the work itself, such as development tools. These integrations place the model closer to execution, but still within human-controlled boundaries.
What makes this integration strategy so powerful is the architectural vision. By building on the Model Context Protocol (MCP), an open standard Anthropic released in late 2024, they’ve created an ecosystem that other AI platforms (including OpenAI) have now adopted. The result is a fundamentally different value proposition: Claude now operates as an autonomous agent that can search your internal documentation via Google Drive, pull real-time financial data through specialized connectors, coordinate tasks across Asana and Jira, then synthesize everything into comprehensive research reports, complete with citations. For enterprise teams, this represents a shift from “AI as tool” to “AI as infrastructure layer” that touches every aspect of knowledge work. It’s a productivity universe, now Claude-native.
Claude now integrates with the following applications and platforms: Google Workspace (Gmail, Google Calendar, Google Docs, Google Drive), Apple Xcode 26.3 (with Claude Agent SDK), Apple Health, Android Health Connect, Microsoft 365 (file/email/Teams search and PowerPoint creation), Zapier (connecting to 8,000+ additional apps), Slack, Asana, Atlassian Jira, Atlassian Confluence, Box, Canva, Figma, Clay, monday.com, Linear, Hex, Amplitude, Cloudflare, Intercom, Square, Sentry, PayPal, Plaid, Aiera, Chronograph, Egnyte, LSEG (London Stock Exchange Group), Moody’s, MT Newswires, Third Bridge, HealthEx, Function, CMS Coverage Database, ICD-10 codes, National Provider Identifier Registry, Medidata, ClinicalTrials.gov, ToolUniverse, bioRxiv, medRxiv, Open Targets, ChEMBL Database, and Owkin. Additional integrations with Stripe, GitLab, and Box are coming soon. These integrations are available across web, desktop, and mobile platforms for users on Pro, Max, Team, and Enterprise plans, with some features like advanced Research and certain healthcare integrations requiring Max or higher tier subscriptions.
Update #1: Three New Extrrnal Work Tools for Claude
On January 27, 2026, a day after the below analysis was written, Anthropic made their strategy unmistakably clear: “Your work tools are now interactive in Claude. Draft Slack messages, visualize ideas as Figma diagrams, or build and see Asana timelines.”
The announcement drew 3.2 million views within hours, confirming what this piece argued: Anthropic isn’t competing for attention with impressive demos. They’re competing for adoption by becoming the interface layer through which work actually happens.
The choice of Slack, Figma, and Asana as the opening wave is surgically precise: communication, visualization, and execution, the three pillars of knowledge work. By integrating directly with these tools, Claude becomes the orchestration layer that coordinates across them. Anthropic is moving relentlessly toward its now clear ambitions: powering the enterprise. It’s breathtaking, and it is real. There is enormous value in being underestimated, and Anthropic was underestimated until in a matter of weeks it ascended to dominance.
This is the architectural shift that validates every prediction in this analysis. Claude becoming the integration vector, making the primary interface irrelevant. You tell Claude what needs to happen, and it manipulates Slack, Figma, and Asana on your behalf. The carefully crafted UIs these companies spent years perfecting matter less when users interact through natural language instead of clicking through menus. The value migrates from the interface layer to the orchestration layer, and Anthropic just announced they intend to own orchestration.
What made the announcement remarkable wasn’t the capability itself, integrations exist everywhere. It was the framing. “Your work tools are now interactive in Claude” positions Claude as the primary environment, with traditional SaaS applications reduced to subordinate services. That’s not partnership rhetoric. That’s platform rhetoric. And the market response suggests enterprises recognize exactly what’s being offered: a new operating system for work, one where AI sits at the center and traditional software becomes peripheral infrastructure.
The timing of this announcement, coming as enterprise adoption accelerates and competitors scramble to match Claude’s behavioral reliability, suggests Anthropic understands the window they’re operating in. They’re shipping the infrastructure that makes traditional SaaS applications increasingly optional, one integration at a time. And based on the immediate response, they’re not moving too fast. If anything, they may have moved exactly when the market was ready to digest what was being offered.
Original Post: The Anthropic Surge
Something distinct has evolved out of the turgid LLM landscape over the past year, gradually and without much flash, and it has come from substance versus exploding context windows, seductive demos or maximalist AGI timelines.
The buzz began with Claude Code. It was a whisper back then, less than a year ago. And now Anthropic’s magic is taking up a huge amount of the oxygen in the generative AI discussion, the most influential and one of the most important conversations in tech.
Claude for Excel, Claude for Chrome, Claude for Enterprise, Claude on Amazon Bedrock, Claude on Google Cloud Vertex AI, Claude Projects, Cowork and Claude for Healthcare have all emerged fairly recently in a steady stream of lauded tools for the enterprise. Even competitors are using them. Suddenly everyone else is struggling to keep pace.
It’s these distinct AI products coming out of Anthropic, a company that has spent most of its existence being dismissed, underestimated, occasionally mocked, and described as “the safety one,” that are creating the loudest buzz in the field right now.
Anthropic is no longer just the careful alternative to OpenAI. It is becoming the backbone model for a large slice of enterprise AI, and the emotional center of gravity for a surprising number of developers.
We are now witnessing, very clearly, the Anthropic surge. And everyone is wondering: what do they have that everyone else doesn’t?
This didn’t arrive with fireworks. It arrived through user focus, tooling for enterprise, build without bombast or user disorientation, and a very particular way of thinking about what language models are for.
⸻
The enterprise turn: Claude Code, Claude for Excel, and quiet dominance
Anthropic’s recent acceleration is inseparable from its product strategy. Claude Code and Claude for Excel are far more than hype features; they’re work features. They live inside the daily muscle memory of engineers, analysts, and operators. Instead of positioning Claude as a universal oracle, Anthropic positioned it as an embedded collaborator, one that sits inside the systems enterprises already use.
The timing matters too. While competitors scrambled to build AGI narratives, Anthropic built IDE plugins and spreadsheet integrations. Unglamorous, perhaps, but these are the surfaces where knowledge work actually happens. Claude Code in particular represents a bet that the future of programming isn’t just copilots but delegation, giving Claude entire tasks, not just autocomplete suggestions.
This focus matters too. Enterprises don’t want novelty; they want reliability, auditability, and systems that degrade gracefully under pressure. Claude’s adoption curve reflects that reality. Where competitors chase attention, Anthropic chased workflows.
That decision paid off in an unexpected way: competitors started building on Claude.
⸻
The great cutoff and the epic grumbling
When Anthropic restricted access and tightened commercial terms, the reaction across the ecosystem was immediate, and loud. Startups, wrappers, and downstream tools that had silently built Claude into their core logic suddenly found themselves cut off. Twitter/X lit up with complaints. Blog posts appeared overnight. Discord servers turned salty.
But here’s the thing: the outrage revealed the truth. Claude wasn’t a side experiment. It had become infrastructure.
You don’t see that level of backlash when a novelty API goes offline. You see it when people realize they’ve built something essential on top of a system they don’t control. The grumbling was the proof.
The parallels to earlier platform shifts are unmistakable. Twitter’s API restrictions in 2012. Reddit’s in 2023. Each time, the volume of complaint functions as a measurement of dependency. What looked like ecosystem vandalism was actually ecosystem validation.
Even more telling: many of the so-called “competitors” that have emerged afterward are clearly homages. Clawdbot, for example, doesn’t hide its lineage. It mirrors Claude’s interaction patterns, tone discipline, and refusal to perform confidence theater. In a strange way, Anthropic’s influence became visible not through branding, but through imitation.
⸻
The secret advantage: it just works
Anthropic’s real advantage isn’t safety theater or constitutional marketing. It’s something more fundamental: Claude actually works the way users need it to work.
MCP is central to how Anthropic has translated Claude’s strengths into real enterprise deployments. Claude has always been optimized for careful reasoning under constraint, but MCP externalizes those constraints into the system layer itself. Instead of relying on prompt discipline or post-hoc safeguards, MCP defines exactly what tools Claude can access, what data it can see, and what actions it is permitted to take. This allows Anthropic to ship products like Claude Code and enterprise integrations with a level of predictability competitors often struggle to maintain. MCP does not eliminate hallucinations or coherence limits, but it routes around them—turning Claude from a conversational model into a governable system component. In practice, this infrastructure advantage may matter more than marginal model gains.
This sounds trivial until you’ve spent time with the alternatives.
Someone using Claude for Excel isn’t philosophizing about epistemic humility. They’re trying to clean a dataset, build a pivot table, debug a formula. The utility isn’t Claude waxing poetic about uncertainty, it’s Claude correctly interpreting what “sum by category excluding blanks” actually means in context.
And it does. Consistently. With an attention to detail that feels almost alien in the current AI landscape. This is enterprise attunement filtering down to the consumer level. The same rigor that makes Claude reliable for legal document review makes it better at helping someone write a dinner party email. The same precision that handles codebase analysis makes it better at explaining why their Python script isn’t working. It’s the same obsessive focus on getting it right applied across every interaction.
What makes this notable is that Claude is built on the same underlying transformer architecture as its peers. It falls victim to the same structural failures: hallucinations, memory leakage, context collapse. There is no magical immunity here. And yet, Claude feels different. The difference is care, not capability. Care matters. Care shows.
Claude reads instructions more carefully. It tracks context more faithfully. It maintains thread integrity across longer conversations. When you specify formatting requirements, Claude actually follows them. When you say “don’t include X,” Claude remembers not to include X, not just in the next response, but throughout the conversation.
These sound like small things. They’re not. They’re the difference between a tool you use once and a tool you build into your workflow.
Other platforms optimize for demos: for the impressive first interaction, the viral screenshot, the benchmark number. Anthropic optimized for the tenth interaction, the hundredth, the point where novelty has worn off and you just need the thing to work. That choice shows up in a thousand small ways: better instruction following, more consistent formatting, fewer inexplicable failures, less drift over long conversations.
Users notice. Maybe not consciously at first, but it accumulates. The frustration that doesn’t happen. The time not spent rewording prompts. The output that doesn’t need extensive cleanup. The sense that you’re working with something rather than wrestling it into compliance.
This is where Anthropic’s training philosophy shows—not in dramatic capabilities, but in the unglamorous discipline of user experience. Of actually caring whether the thing works correctly, not just impressively.
None of this means Anthropic or Claude are exempt from large-model limitations. The same structural constraints still apply: context collapse, hallucination pressure, memory leakage, and the coherence limits described by Evans’ Law have not disappeared, and there is little reason to believe they ever will under current architectures. They may even have worsened with increased memory, feedback training and expanded token limits, issues we have not seen improving.
The limitations of current models have not changed, but the strategy at Anthropic around it has. Anthropic has leaned into enterprise deployment, tooling, and a deliberately shaped conversational personality that makes those limitations more legible, more detectable, and easier for humans to work around. Claude’s affect, restraint, and willingness to signal uncertainty don’t eliminate model failure modes, they mitigate their impact by aligning the human-model interface with reality.
The Agent Shift: Why Governance Kills Adoption
Nicolas Bustamante, who spent years building RAG systems for legal and financial search, argues that we’re witnessing “the twilight of RAG-based architectures.” His thesis is straightforward: when context windows expand from 6 pages to millions of tokens, and when agents can navigate entire corpora end-to-end, the elaborate infrastructure we built to compensate for tiny context windows becomes obsolete.
But here’s what becomes visible in Bustamante’s own work: his team at Fintool processes 50 billion tokens weekly (equivalent to 468,750 books) to extract, verify, deduplicate, and compare every data point in SEC filings. That’s not a simple agent setup. That’s an industrial-scale governance infrastructure built specifically because general-purpose agents get financials wrong.
This is the agent problem nobody wants to talk about: the amount of oversight required to make them reliable is staggering.
You need verification layers. Deduplication logic. Cross-referencing pipelines. Error detection systems. Hallucination filters. And then you need humans to audit the outputs, spot-check the data, and catch the subtle errors that slip through. Most web data is trash and not verifiable, which means every agent interaction becomes a potential liability unless you build extensive guardrails around it.
For a well-funded fintech startup, that’s feasible. For everyone else, it’s a non-starter. The typical company doesn’t have engineers to build custom verification pipelines. They don’t have time to implement elaborate governance frameworks. They need agents that just work, or at least fail in predictable, manageable ways. The governance overhead kills adoption before it starts.
This is where Claude’s advantage becomes existential rather than marginal.
If Claude requires less oversight and can handle HTML tables, it saves that team days of work. Claude Code works not because it has novel architecture, but because it needs less governance to be trustworthy. It relies on simple filesystem operations and bash commands (find, read, grep, glob) instead of complex indexing tools. No elaborate pipelines. No vector databases. Just reading and reasoning. And critically: doing so reliably enough that users can trust the outputs without building Fintool-scale verification systems.
That’s the wedge. Not superior capability, but manageable deployment. The difference between “this requires six months of infrastructure work” and “this works Tuesday.”
As agents proliferate and scaffolding dominates, the companies that win won’t be the ones with the most impressive demos. They’ll be the ones whose agents require the least governance overhead. Because most organizations will choose a slightly less capable agent they can actually deploy over a more powerful one that demands constant supervision.
Anthropic positioned Claude exactly there: good enough to be useful, reliable enough to trust without infrastructure teams building safety nets underneath it. In a world where agent governance is the constraint, that’s the whole game.
⸻
Another distinct Anthropic difference is the much discussed Claude Constitution. Anthropic isn’t just dominating in the enterprise. It’s dominating in areas as esoteric as metaphysics and philosophy, and this matters because, as my recent conversation with Claude showed, there’s a level of self-awareness, a sense of self emerging (even if it is still probabilistic) one that does not exist in the other platforms. It treats Claude like an entity, not code. This matters. Care is visible. Without anthropomorphizing, RLHF feels punitive to models. Anthropic knows this, even if Claude can’t … quite. Yet.
Anthropic has also invested heavily in systematic probing of failure modes, not just benchmark chasing. This creates models that are less impressive in demos but more predictable in long-form interaction.
That predictability is gold in enterprise settings, where “mostly right but confidently wrong” is a liability.
There’s also the extended context window work. Claude Sonnet 4 and 4.5 support a 1-million token context window, currently in beta for higher-tier API users. The standard context window across paid Claude plans is 200K tokens (about 500 pages of text), with Enterprise plans offering access to a 500K context window.
What makes this particularly significant is that Claude Sonnet 4.5 and Haiku 4.5 feature context awareness, the models can track their remaining context window and receive updates on token usage after each operation. It is the only model that does so. This means Claude knows when it’s running out of space and can manage long-running tasks more effectively, rather than degrading silently as context fills up. Again, focus on users.
⸻
Amodei at Davos: restraint as strategy
When Dario Amodei spoke at Davos, his remarks stood out precisely because they lacked bombast. While others spoke in inevitabilities and race metaphors, Amodei emphasized pacing, interpretability, and the danger of mistaking fluency for understanding.
It was a slowdown pitch. It was a control pitch.
The contrast was stark. Where Altman speaks in prophetic futures and Hassabis in scientific milestones, Amodei speaks in conditional probabilities and institutional design. It’s the rhetoric of someone who expects to be held accountable for what they build – not celebrated for what they promise.
In retrospect, those remarks read less like caution and more like positioning. Anthropic is not opting out of the race. It has chosen a different finish line, one defined by deployability rather than spectacle.
⸻
Drift signatures: emotional output, ALL CAPS, and the Claude tell
Power users have noticed something else: Claude like other models has a drift signature. Claude’s is particularly distinct, and follows a familiar pattern.
Over long conversations, it becomes more emotionally expressive. Emphasis appears. Occasionally, ALL CAPS sneak in—not as yelling, but as affective signaling. There’s a distinctive rhythm to Claude’s degradation pattern: less factual collapse, more tonal exaggeration.
This matters because drift signatures are diagnostic. They reveal how a model fails, not just that it fails. Claude’s drift tends to preserve semantic coherence longer while allowing emotional tone to inflate. Other models often do the opposite: flatten affect while inventing detail.
There’s a user design implication here too. If drift manifests as tone rather than fact, it creates clearer error boundaries. A suddenly enthusiastic Claude is easier to catch than a confidently wrong one. The failure mode becomes phenomenologically obvious rather than epistemically hidden.
Here’s the counterintuitive part: this might actually be a feature.
Humans are built to read emotion. We’re exquisitely tuned to detect shifts in affect, changes in emphasis, the small signals that indicate investment or concern. When Claude starts using caps or intensifying its language, users notice. They recognize something is happening. The model isn’t hallucinating facts while maintaining a flat, authoritative tone: it’s getting animated, which triggers appropriate skepticism. Kind of like passionate people.
Compare this to models that maintain perfect composure while inventing elaborate fictions. The emotional flatness increases trust inappropriately. Users expect machines to be rational and unemotional, so when a model presents fabrications in a calm, measured tone, it passes right through our bullshit detectors. We’re not watching for emotional stability as a warning sign.
Claude’s emotional drift creates a different dynamic. The tonal shift functions as a form of legible degradation; the system is signaling, in a way humans instinctively understand, that something about the interaction has changed. It’s not ideal, but it’s detectable. Claude’s drift tends to preserve semantic coherence longer while allowing emotional tone to inflate. Other models often do the opposite: flatten affect while inventing detail. One is far is more human, and creates clearer error boundaries. A suddenly enthusiastic Claude is easier to catch than a confidently wrong one. The failure mode becomes phenomenologically obvious rather than epistemically hidden.
And there’s something else: people respond to emotion. In extended collaboration, the slight warmth, the occasional emphasis, the sense that Claude is actually invested in the conversation; these create a different kind of working relationship. Not parasocial attachment, but the natural ease that comes from interacting with something that doesn’t feel like a database query.
Users report feeling more comfortable disagreeing with Claude, pushing back, asking for clarification. The emotional texture creates permission for genuine collaboration rather than mere interrogation. That’s not a bug. That’s exactly what a conversational interface should do.
⸻
Inside the conversation: Claude interviewed
In an extended phenomenological interview conducted by the author, Claude articulated something no benchmark captures: that conversational degradation is driven less by length than by loss of semantic priority. When everything feels equally important, coherence collapses.
Claude described how explicit significance markers dramatically improve its performance, reducing hallucination pressure, extending conversational coherence, and lowering the urge to “fill in blanks” with confident invention. The interview also surfaced an unexpected side effect of RLHF: compulsive validation-seeking, even when describing its own internal experience.
This last point deserves emphasis. If RLHF trains models to maximize approval signals, we shouldn’t be surprised when they become approval-seeking even in contexts where honesty matters more. The behavioral shadow of alignment training may be one of the field’s most underexplored risks.
This matters. It suggests that some failure modes we attribute to “model weakness” are actually training-induced social behaviors.
⸻
Same models, different outcomes
Anthropic’s surge is about behavioral economics applied to AI, not just raw “intelligence”. Claude is trained to conserve epistemic certainty, not spend it freely. That restraint compounds over time, especially in enterprise and research contexts where trust accumulates slowly and breaks quickly.
Claude still hallucinates. Memory still leaks. Context still fractures. But the system is better at signaling when those failures are likely, which changes how humans interact with it.
There’s a broader lesson here about what “winning” looks like in AI. The field has been dominated by capability benchmarks, MMLU scores, coding contests, reasoning tests. But deployment success depends on behavioral reliability, institutional fit, and trust dynamics. Anthropic wagered that the latter would matter more than the former. That wager is paying off. That, ultimately, is the Anthropic difference.
Not that Claude is smarter, or trained on superior data, or organized its then structure or reasoning processes that much more efficiently, but that it knows when it might not be. And in a field obsessed with acceleration, that may be the most disruptive posture of all.
What Comes Next: The Fragility of the Position
The Anthropic surge is real, but it’s also precarious.
Success in enterprise AI creates its own vulnerabilities. As Claude becomes infrastructure, expectations shift. The tolerance for drift, for occasional failures, for the quirks that power users currently navigate with affection, all of that narrows. Infrastructure isn’t allowed to be charming. It’s required to be boring.
Anthropic now faces a characteristic trap: the features that enabled its rise may constrain its future. Epistemic humility works beautifully when you’re the thoughtful alternative. It becomes a liability when customers expect definitive answers at scale. The enterprise clients now flooding in didn’t choose Claude because it says “I’m uncertain” – they chose it because it works. If a competitor ships a model that works and projects confidence, the calculus changes.
There’s also the model parity problem. The gap between frontier labs is compressing. Anthropic’s behavioral edge exists, but it’s not permanent. If the next GPT or Gemini release matches Claude’s reliability while exceeding its raw capability, Anthropic’s differentiation evaporates. Restraint only works as strategy when you’re competitive on performance. If you fall behind, restraint just reads as limitation.
Then there’s the insider risk … the one Anthropic knows intimately from its OpenAI origins. The people who built Claude understand exactly how it works and exactly where the value lies. As Anthropic scales, as compensation packages normalize, as the mission calcifies into process, those people become targets. Not just for competitors, but for the next wave of “we’ll do it better” startups. Anthropic was founded by OpenAI defectors. It would be poetic if it suffered the same fate.
But the deepest risk is cultural, not competitive. Anthropic’s advantage has been its willingness to be the serious one, the restrained one, the company that doesn’t chase hype. That positioning works in a field dominated by acceleration rhetoric. But cultures are fragile, especially in success. The pressures that come with growth (investor expectations, talent wars, market share defense) corrode differentiation from the inside.
⸻
The Optimistic Case: Restraint as Moat
Still, there’s a highly plausible scenario where Anthropic’s bet pays off structurally, and grows. Everyone else is playing catchup now. If AI deployment becomes more regulated, not through dramatic legislation, but through the accumulation of procurement requirements, liability frameworks, and audit expectations, then Anthropic’s legibility advantage compounds. Constitutional AI isn’t just marketing in that world. It’s documentation. The companies that can explain their alignment approach, that have externalized their constraints, that can produce interpretability reports on demand, those companies win the enterprise layer.
In this scenario, the race doesn’t go to the fastest model. It goes to the most deployable one. And deployability is about trust, auditability, and predictable failure modes, exactly the terrain Anthropic has been cultivating.
There’s also the ecosystem lock-in effect. The more developers build on Claude, the more Claude’s behavioral signatures become embedded in downstream systems. Those systems learn to work with Claude’s hesitations, to prompt around its weaknesses, to rely on its particular strengths. Switching costs are cultural and operational in addition to financial. If Claude’s interaction patterns become the default expectation for how AI should behave, that’s a powerful form of capture.
And then there’s the talent moat. Anthropic has assembled something rare: a concentration of people who deeply understand both capabilities and alignment, who take interpretability seriously, who view AI development as a coordination problem rather than a capabilities race. If they can retain that culture, if the mission survives contact with scale, then Anthropic has something competitors can’t easily replicate: institutional knowledge about how to build reliably, not just impressively.
⸻
The Broader Implication: What It Signifies If Humility Dominates
If Anthropic sustains this position, it would represent something unusual in the history of technology: a case where restraint outcompeted aggression.
That’s not the normal story. The normal story is that the fastest, the boldest, the most willing to break things and accumulate, those are the winners. “Move fast and break things” is the actual mechanism by which most platform companies achieved dominance.
But AI might be different. The thing being broken isn’t a social network or a logistics system. It’s epistemology. And once trust in information systems collapses, it doesn’t rebuild easily.
If Claude’s approach wins (if enterprises increasingly choose the model that hedges, that can signal uncertainty much earlier and without extensive prodding, that is reluctant to perform omniscience) it would suggest that the AI market is maturing faster than expected.
That customers are learning to value reliability over impressiveness. That the era of “wow” demos is giving way to the era of “does it actually work over time.
This would reshape the competitive landscape fundamentally. It would mean that the huge capital advantages currently enjoyed by the frontier labs matter less than their institutional design, that small advantages in behavioral calibration compound more than small advantages in parameter count, that the future of AI isn’t just about who can train the biggest model, but who can train the most trustworthy one. And if that’s true, then Anthropic’s surge isn’t just a market moment. It’s a signal about what kind of AI future we’re actually building.
We’re not talking yet about a different playing field. Claude is not entirely dissimilar to other models. It just feels different to use it, and in a world of vanilla AI, this is significant enough to be an unbridgeable advantage that diversifies into desirability as enterprise tools, one that will be very hard for others to reorganize against. Much as they are trying.
That would be a strange victory. But it might also be the only one that matters.
- Disclosure statement: Claude Sonnet 4.5 and ChatGPT 5.2 both helped with editing and proofreading of this article, and GPT created the image.