Before an agent gets more capable, it needs somewhere safe to stop

Connecting a model to search, databases, and a code executor proves that it can call tools. It does not prove that it can run unattended. Production architecture begins with authority, stopping conditions, and recovery.

Map the workflow first

Map one task into explicit states: accept input, plan, call a tool, verify the result, and return an answer. Give every state an allowed transition, timeout, and retry limit. Anthropic similarly recommends simple, composable workflows before autonomous agents.

Keep authority separate from capability

Grant the minimum tool permissions required by the task. Separate reads from writes, require confirmation for messages, payments, deletion, and deployment, and keep credentials in the tool layer rather than the prompt.

Give failure a deterministic exit

Set task-wide limits for time, steps, and spend. Retry transient failures with backoff, stop on permission or validation errors, and escalate uncertain results. Never let the model repeat the same failed call indefinitely.

Pre-release checks

• A staging task can be replayed end to end
• Every side effect can be blocked or reversed
• Budget, step, and time limits produce an explicit failure

Sources

• Anthropic: Building effective agents
• OpenTelemetry GenAI conventions