Anthropic expands Claude Managed Agents API with effort levels, lifecycle webhooks, and session seeding

Anthropic shipped a batch of Claude Managed Agents API updates on July 22, giving developers finer control over agent cost, faster session startup, and webhook-driven visibility into agent infrastructure instead of polling.

What's new

Four changes landed together in the Claude Platform release notes:

• Per-agent effort levels. Developers can now set an effort level directly inside the model object when creating a Managed Agents agent, letting teams tune reasoning depth (and cost) per agent rather than relying on a single account-wide default.
• Lifecycle webhooks for environments and memory stores. Webhooks now cover four environment.* event types and three memory_store.* event types. As Anthropic put it, "You can react to environment and memory store lifecycle changes without polling."
• Session seeding. New Managed Agents sessions can be created with initial_events on POST /v1/sessions — up to 50 user.message and user.define_outcome events. A non-empty list starts the agent loop in the same call, removing the need for a separate follow-up request to kick off work.
• Optional version field on agent updates. The version field is no longer required when updating a Managed Agents agent. Supplying it enables optimistic concurrency (a mismatch now returns a 409 error); omitting it applies the update unconditionally.
• Thread-level event deltas. Session thread event streams now support event_deltas, so GET /v1/sessions/{session_id}/threads/{thread_id}/stream can preview a subagent's text as the model generates it, mirroring the delta streaming already available at the session level.

Context

These updates build on a steady cadence of Managed Agents platform work Anthropic has shipped through June and July: event-delta streaming at the session level (June 30), agent/deployment lifecycle webhooks (June 30), per-session configuration overrides (June 30), and Dreams research-preview support for Claude Fable 5 and Sonnet 5 (July 10). Managed Agents is Anthropic's hosted infrastructure for running Claude-based agents in production — handling session state, memory, tool execution, and now finer-grained observability — rather than developers stitching that infrastructure together themselves on the raw Messages API.

Why it matters

None of these changes are flashy on their own, but together they target the two biggest complaints from teams running agents in production: cost control and operational visibility. Per-agent effort levels mean a company running dozens of specialized agents no longer has to pay premium reasoning costs for simple ones. Lifecycle webhooks and thread-level deltas replace polling loops — a real latency and API-quota cost at scale — with push-based updates. And session seeding cuts one network round-trip off every agent invocation, which compounds quickly for high-volume deployments. It's the kind of incremental platform hardening that signals Managed Agents is moving from beta convenience features toward the reliability and cost controls that production deployments actually require.