On March 23, 2026, computer use in Cowork and Claude Code was still a research preview.
That label is a useful engineering constraint: a capable screen-driving agent is not automatically a production-ready system.
Start with the integration boundary
The safest default is a structured integration. APIs usually give engineers better reliability, speed, testability, and cost control, so they should handle every step they can cover cleanly.
Visual control belongs in the remaining gap. It is useful when a workflow is GUI-only, trapped in legacy software, blocked by an incomplete API, or spread across applications. That boundary matters because screen control introduces a less structured execution surface.
Map the workflow before choosing the agent. Separate steps with stable API coverage from steps where the interface is the only practical path. A hybrid design can keep deterministic work structured while reserving visual action for the narrow part that needs it.
Put controls around the task
A prompt is not a production control. The operating environment needs to constrain what the agent can see, change, and continue doing.
Before rollout, require:
- Isolation from unrelated applications, files, and credentials.
- Least-privilege access limited to the task.
- Approval gates before consequential actions.
- Stop conditions for unexpected states or repeated failure.
- Complete action records that show what the agent attempted.
- A defined recovery path when the run does not finish safely.
These controls have an honest cost. Approval gates reduce uninterrupted autonomy, while tight permissions can block an unforeseen but legitimate step. The alternative is an agent with a wider blast radius and less oversight. Production design is the act of choosing that boundary deliberately.
Evaluate the run, not the demo
A polished screen recording answers only whether one path worked once. Task-level evaluation has to judge the final state and the behavior that produced it.
Final-state correctness should be explicit: which file, field, or application state must exist when the task ends? Then measure retries, latency, token budget, unsafe attempts, and recovery. A run that eventually reaches the right result after unsafe actions should not pass merely because the last screen looks correct.
Retries expose brittle navigation. Latency shows whether the visual path is operationally acceptable. Token budget keeps the evaluation connected to cost control. Unsafe-attempt tracking tests whether the controls actually intervene. Recovery checks whether the workflow can return to a known state instead of leaving partial changes behind.
The test set should include the states that trigger an approval, a stop, and a recovery. That makes the safety model observable rather than aspirational.
Read product status precisely
The timeline shows why capability and readiness should be evaluated separately. Anthropic introduced computer use with Claude 3.5 Sonnet in October 2024 and opened it through the Anthropic API. On March 23, 2026, release notes described direct computer use in Cowork and Claude Code as a research preview. Cowork itself reached general availability on April 9, 2026.
Those labels can coexist because a generally available product may contain a capability with a different maturity designation. Engineers still need to validate the exact workflow, controls, and failure behavior they plan to operate.
CNBC's March 24, 2026 coverage described Claude navigating applications, working with spreadsheets, and completing multi-step desktop tasks. That demonstrates a broader task surface; it does not remove the need for application-level boundaries.
Make the rollout decision
Prefer an API when it can complete the step. Add visual control only when the interface is the practical route. Isolate that route, grant the minimum access, place approvals before consequential actions, define stop conditions, and preserve the full action record.
Then evaluate correctness, retries, latency, token use, unsafe attempts, and recovery as first-class outcomes. If the team cannot state what stops the agent or how a partial run is recovered, the workflow is not ready for production.
Which step in your current workflow truly requires visual control, and what approval would you place immediately before it?
📖 Read the full guide → Computer-Use Agents: An Engineer's Production Guide
Top comments (0)