Anthropic pushed Computer Use, the Skills API, and the Files API to general availability on the same day, with Browser Use joining as a new addition. Four tools in one drop, and together they cover every angle of a complete agent workflow.
Anyone who has actually shipped Claude into a real business knows the pain: a model can be brilliant and individual capabilities can be flashy, but the moment one piece is missing, nothing runs end-to-end. What Anthropic is targeting this time isn't "wow a demo," it's "an agent that independently delivers a finished task."
Four Tools, One Closed Loop
Stripped down, the four tools map to four actions:
- Read: Files API ingests context
- Load: Skills inject domain knowledge on demand
- Act: Computer Use / Browser Use operate external systems
- Write: Files API writes results back
The biggest headache for anyone running production agents is fragmented state. Files sitting in S3. Knowledge buried in Confluence. Web pages that have to be scraped from scratch. Screenshots fed back into the model. This suite collapses the whole chain onto one tool stack. Files, memory, actions, and outputs all hang off the Files API, giving the agent something like its own working disk.
Why does this matter? Because agents that look magical in demos fall apart in production. Uploading a file meant re-encoding it in base64. Every session started from zero. Knowledge could never fit inside the system prompt's token budget. Files + Skills effectively separates "the conversation" from "the persistent state."
Computer Use's Real Killer Feature: Single-Turn Multi-Action
The biggest change in Computer Use GA isn't "it works now" (it worked in beta). It's single-turn multi-action: a single model response can click, type, and submit in sequence without waiting for the model to come back between each action.
Before, the loop was "click → wait for model → click again," with latency piled up on every model round-trip. Now a single response packs in multiple actions, and the loop keeps going as the model continues to emit action sequences. Anthropic disclosed one example: a medical claims customer compressed its longest workflow from 32 minutes to 13 minutes, dropped per-task cost by roughly 30%, and hit a 100% completion rate.
I've run similar scenarios inside Claude Code myself. Before optimization, a form with five fields meant five round-trips, and at 2–3 seconds per round-trip, that was 15 seconds of dead waiting. With multi-action, that latency basically disappears. It feels like a different product.
Also landing alongside: HIPAA compliance support. Computer Use can now operate inside healthcare back-office systems and stay compliant at the same time. That wasn't even on the table before. Claims processing, EMR workflows, patient intake — these are now realistic workloads to hand off to an agent.
Browser Use Stops Guessing at Coordinates: DOM + Vision Is the Answer
The early Computer Use approach was pure vision: take a screenshot, click on pixel coordinates. It looked spectacular in demo videos and collapsed in production. A page redesign, a banner ad popping up, a font size changing — shift 5 pixels and the model clicks "cancel order" instead of "delete order."
The core change in Browser Use is the introduction of dual-channel localization with DOM structure + vision (per Anthropic). It reads the DOM first to grab element IDs, classes, and text, then layers the vision model on top to confirm position. The two signals validate each other: if the element is stable, lean on the DOM (fast, reliable, verifiable). If the element moves, fall back to vision (fault-tolerant).
The official framing is "more stable, better suited for production." From an engineering angle: the DOM signal answers "what to click," and the vision signal answers "where to click." Decoupling them means each can evolve independently. For production agent teams, this is huge. Every frontend release used to be a gamble about whether the agent would go on strike. Now there's a stable anchor in the DOM.
Skills API: Anthropic's Container for Enterprise Knowledge
Skills is not another prompt template marketplace. Anthropic defines it as: versioned Skill units, loaded on demand, entering context only when needed.
A quick comparison against common existing approaches:
- Hardcoded knowledge in the system prompt: eats tokens and pollutes every conversation.
- RAG retrieval: good for documents, awkward for SOPs and runbooks.
- Tool-call documentation: scattered across tool descriptions, hard to maintain.
- Skills: "this is a capability, invoke when needed," and it's versionable.
A concrete example: a medical claims Skill bundle might include HIPAA-compliant field mappings, ICD-10 coding rules, and templates for calling insurance carrier APIs. When the agent takes on a new case, it loads that Skill on demand instead of stuffing the whole manual into context every time.
This is essentially pushing Claude's capability boundary from the prompt layer down to the filesystem layer. For enterprise users, the upshot is that knowledge finally has a proper home — manageable like code, with branches, versions, and staged rollouts.
Files API by the Numbers: 5x Rate, 1 TB, 500 MB
A few key numbers from Files API GA (per Anthropic):
- Rate limits raised 5x
- 1 TB storage per organization
- 500 MB per-file cap
- Automatic expiration from 1 hour to 90 days
For long-running agents, "upload once, reference many times" is a hard requirement. Previously, every session re-uploaded files, and base64 encoding alone ate several seconds of round-trip time. In multi-turn conversations, when the agent wants to look back at a historical file, Files API hands back a reference and skips the retransmit.
Automatic expiration matters too. Production agents regularly run scheduled jobs (end-of-day reconciliation, weekly report generation), and an expiration policy removes the manual cleanup burden while satisfying compliance requirements.
Anthropic Wrote Down the Risks Themselves: Sandboxes, Domain Allowlists, Least Privilege, Human Confirmation
Browser Use + Computer Use can operate real browsers and real desktops, and Anthropic didn't soft-pedal the side effects. The risk list is explicit:
- Prompt injection: malicious instructions on web pages hijacking the agent
- Out-of-scope access: the agent wandering outside its intended boundaries
- Data leakage: screenshots or files escaping the sandbox
- Misclicks: hitting the wrong button and causing business damage
Four defensive postures the team recommends:
- Sandboxed containers: agents run inside Docker or VMs, never touching production
- Domain allowlists: only specific domains are reachable — no free-roaming the web
- Least privilege: API keys and database credentials scoped strictly to what's needed
- Human confirmation: pop-up confirmations before high-stakes actions (Claude now supports a confirmation mode)
My take: an agent that can operate a browser isn't "ready out of the box" for production. It needs to be caged. Anthropic admits as much — the risks document runs a full page.
A Quick Note: Claude Code v2.1.237
Claude Code shipped v2.1.237 the same day. Two changes worth flagging:
- Official Concise output style: agents used to narrate "I am going to do X…" "I am now doing Y…" — pure noise. The new style outputs action commands directly, saving tokens and cutting chatter.
- Prompt caching support fixed for LLM gateways: routing Claude through gateways like LiteLLM or Portkey used to tank cache hit rates. After this fix, enterprise users going through gateways can finally realize the cost benefits of caching.
Neither change is huge, but for teams deploying Claude Code in an enterprise setting they're genuinely useful: one saves tokens, the other saves money.
A Personal Take
With these four tools in GA, the infrastructure for building agents is mostly there. The next competitive frontier isn't "can you build it" — it's "can you run it stably."
- Observability: an agent ran for 13 minutes — was every step on the expected path? Who audits it?
- Cost modeling: tokens consumed per task, file storage used, Skills loaded how often — these numbers are still mostly black boxes
- Evaluation frameworks: per-task completion rate isn't enough. The real questions are long-run stability, SLAs, and failure modes
Anthropic's wave here solved "agents can do work." The next wave has to solve "agents can do work reliably." The latter is harder.
Reference: Anthropic Agent Tools GA announcement