Choose a managed runtime or control the loop yourself
The Messages API fits an application that controls each step itself. Managed Agents supplies the agent loop and runtime. Research across several documents with a reviewable result file is a plausible candidate. A single classification inside an existing automation often does not need this additional state management.
Separate the agent, environment and individual task
An agent defines behavior and tools, an environment defines execution location, and a session represents the individual task. This avoids creating uncontrolled configurations for each customer request. Werkverstand would version configurations and let the application select approved versions. Each task also needs an identifier, a budget and a designated result location.
Make application approval explicit
Managed permission policies govern server-executed agent and MCP tools; custom tools executed by your application remain its responsibility. For a CRM draft, that means narrow read access and a separate write decision. A repeated event must not create the same contact twice. Idempotency therefore belongs in the interface, not only in a prompt.
Set retention rules before using real data
Documentation currently excludes Managed Agents from Zero Data Retention and HIPAA BAA coverage. Sessions retain state and outputs, while uploaded files need separate treatment. Use cleaned data in the pilot. Subsequent approval should identify data categories, deletion, access and treatment of intermediate results.
Prove completion with the actual result
A green event alone does not prove the work is correct. The application checks that the result file exists, meets its schema and passes domain criteria. Also test cancellation, network failures, duplicate delivery and revoked permissions. Only then is it clear whether the runtime actually reduces operational work.
Keep it verifiable
Primary sources
- Claude Managed Agents overviewSource checked:
- Managed Agents permission policiesSource checked:
