Anthropic announced on 20 August 2026 that computer use, the Skills API, and the Files API are generally available on the Claude Platform. The release also introduces a browser-use tool and an updated computer-use tool that can execute multiple actions in one turn. For enterprise teams, the material change is not simply wider access to agent features. It is the need to govern a larger action, instruction, and data surface as these capabilities move into routine deployment.
General availability is a vendor lifecycle milestone, not evidence that a capability is approved for every organization or workload. Computer and browser agents can act in user interfaces, skills can package reusable instructions and code, and files can carry business data across workflows. Each layer needs an owner, an authorization boundary, a change process, and auditable evidence before production use.
What Anthropic released
According to Anthropic's announcement, computer use, the Skills API, and the Files API have moved to general availability on the Claude Platform. The company also released a browser-use tool and updated computer use to support multiple actions within one model turn. Anthropic says this can reduce latency for multi-step tasks, but any performance result remains a vendor claim that should be tested in the buyer's own environment.
The Skills API lets developers upload and version reusable packages that give Claude specialized procedures and resources. Anthropic says the Files API now supports automatic expiration, rate limits five times higher than before, and up to one terabyte of storage per organization. The announcement says Skills and Files are also available through Microsoft Foundry. It says updated computer and browser use are planned for Google Cloud Vertex AI, so teams should treat that part as future availability rather than a current cross-cloud capability.
These services connect three distinct control planes:
| Capability | Operational value | Governance question |
|---|---|---|
| Computer and browser use | Acts across graphical or web interfaces | Which actions, accounts, and destinations are permitted? |
| Skills API | Reuses instructions, code, and resources | Who approved this version and what changed? |
| Files API | Stores and reuses workflow data | What data class, retention rule, region, and access policy apply? |
Why general availability raises an agent-governance decision
A conventional model endpoint returns content. A computer-use agent can navigate interfaces and take actions, which means an incorrect or manipulated output can have an external consequence. Microsoft's guidance for browser automation agents independently warns that such agents may access credentials and enterprise systems, make mistakes, or be misled by malicious data. Microsoft places responsibility on the deployer and recommends sandboxing, minimum privileges, careful monitoring, and human oversight for sensitive actions.
The same reasoning applies to reusable skills. A skill can improve consistency, but it can also distribute outdated instructions, unsafe code, or permissions to many workflows. Versioning is useful only if an organization records which version ran, who reviewed it, which dependencies it used, and when it must be reassessed. Microsoft's Foundry documentation similarly describes skills as reusable packages and makes their execution environment and lifecycle part of the deployment design.
Files create a third governance obligation. Automatic expiration can support minimization, but a default expiry is not a legal retention policy. Teams still need to classify the data, establish purpose and residency requirements, restrict retrieval, delete derived copies, and document exceptions.
Maetra's guide to enforcing AI governance policies in production explains why a policy must connect to runtime decisions. The Task Guard overview shows how intended task, observed action, and escalation evidence can be compared for autonomous workflows.
A deployment checklist for computer use, skills, and files
AI-platform, security, compliance, and product owners should require eight items before enabling these capabilities for a production workflow:
- Name the workflow and owner. Document the business purpose, users, data, environments, expected actions, prohibited actions, and accountable service owner.
- Define the action envelope. Allow only the applications, domains, accounts, tools, and transaction types required. Require separate approval for messages, submissions, purchases, access changes, or destructive actions.
- Use isolated identities. Run agents in a sandbox with short-lived credentials and least privilege. Do not let a general browser session inherit a human administrator's full account access.
- Govern skill versions. Review source, dependencies, permissions, network behavior, expected output, and rollback procedure. Pin the approved version and trigger reassessment when any material component changes.
- Set file rules. Classify uploads, configure access and expiry, verify location and deletion requirements, and prevent a file from crossing tenant, matter, or privilege boundaries.
- Keep humans at consequential boundaries. Require an informed reviewer before legal commitments, regulated decisions, external publication, money movement, high-impact communications, or irreversible changes.
- Log for reconstruction. Capture the model, policy and skill versions, input classification, tool calls, destinations, authorization results, file identifiers, approvals, errors, and final outcome.
- Test failure and recovery. Evaluate prompt injection, malicious webpages, ambiguous instructions, stale skills, unexpected dialogs, repeated loops, and credential exposure. Provide a kill switch and a tested incident procedure.
Procurement should also verify service terms, subprocessor arrangements, regional availability, usage limits, support commitments, and the exact controls available in the chosen cloud route. A feature having the same name across platforms does not prove identical behavior or configuration.
Limits and the next control decision
The release is Anthropic's account of its own product. General availability does not establish security, compliance, or fitness for a particular purpose, and vendor performance claims require independent testing. Browser content can be hostile, graphical interfaces can change, and an agent can misunderstand a valid instruction. Cloud availability also differs: the announcement describes the Vertex AI update as coming later.
The immediate decision is therefore not whether every team should adopt the tools. It is which workflows have a defensible action envelope, isolated credentials, governed skill versions, controlled files, and sufficient evidence for approval. Teams without those elements should keep the capability in a restricted test environment.
Maetra analysis: agent release gates should bind approved purpose to actual actions, skill versions, file access, and human escalation. Maetra Govern can help teams assign those controls and evidence requirements before broad deployment.