Claude's Agents API adds live subagent streams and session seeding
Anthropic's quiet dev-platform drop also adds lifecycle webhooks to kill polling, per-agent effort dials, reverse pagination, and scoped credentials
Copy markdown
Seed a whole session in one POST
POST /v1/sessions now takes an initial_events array — up to 50 user messages or defined outcomes — and kicks off the agent loop immediately, so you skip the separate send-events round trip when you spin an agent up.
Stream a subagent's text as it's written
The thread event stream now emits event deltas, so a multi-agent UI can preview a subagent's output as the model generates it instead of waiting for the finished turn.
Lifecycle webhooks, so you can stop polling
New environment.* and memory_store.* webhook event types fire on infra state changes — react to them instead of hammering the API on a poll loop.
An effort dial per agent, plus per-session overrides
Set an effort level right on a Managed Agent's model config, and override it per session — so a cheap triage agent and a max-effort planner can share one deployment.
Reverse pagination and scoped credential injection
Two quality-of-life wins: page threads newest-first, and scope exactly which injected credentials each agent sees. Agent updates also take an optional version field for optimistic-concurrency safety.