Claude's agent API adds effort control, session seeding, live deltas
Anthropic's Managed Agents API can now tune reasoning depth, preload up to 50 events per session, stream token deltas, and fire lifecycle webhooks.
Copy markdown
Dial an agent's reasoning up or down
You can now set an `effort` level in a Managed Agent's `model` config at creation, tuning how hard it thinks per task — trade latency and cost against depth without switching models.
Warm-start a session with 50 events
New `initial_events` on `POST /v1/sessions` seeds a session with up to 50 `user.message` and `user.define_outcome` events, so an agent can start mid-context instead of replaying the whole history.
Stream token-level previews
Session thread streams add an `event_deltas[]` parameter for real-time text preview as the agent generates — the missing piece for a live typing UI on top of Managed Agents.
Webhooks for environments and memory
Four new `environment.*` and three `memory_store.*` webhook event types fire on lifecycle changes, so you can react to environment spin-up, teardown, and memory writes instead of polling for them.
Push agent updates without a version
The `version` field is now optional when updating a Managed Agent: omit it to force the change, or pass it for optimistic concurrency that 409s on a mismatch — handy for CI that ships agent configs.