Claude Managed Agents adds per-agent effort controls and 500 skills per session

• Claude Managed Agents now exposes per-agent effort, seeded session creation, webhooks, sub-agent event streaming, and a 500-skill session ceiling, according to ClaudeDevs' launch post.
• Effort became a cost and latency knob, with ClaudeDevs on effort saying lower settings give faster, cheaper responses.
• Session startup lost an API round trip: ClaudeDevs on seeded sessions says a create call can now include up to 50 user_message and define_outcome events.
• Skills got a much larger ceiling, since ClaudeDevs on skills says a session can attach up to 500 across all Managed Agents.
• The app-side pattern is moving toward teach-by-demonstration, with shannholmberg's marketer workflow and ozansihay's Cowork note both describing screen-recorded skills for repeatable desktop and browser work.

Anthropic's Platform release notes include a quiet agent-update change too: version is now optional, so teams can use it as an optimistic concurrency key or omit it for unconditional updates. The sessions docs say initial_events are not echoed on the create response, which is exactly the kind of tiny migration gotcha that eats an afternoon. The cookbook notebook demos a 7th-grade science curriculum team with a high-effort standards researcher, a no-web lesson writer, and live per-thread sub-agent previews.

Anthropic's Managed Agents overview defines the surface as a hosted agent harness for long-running, asynchronous work, with server-side event history, sandbox state, files, tools, and compaction. The July 22 update turns more of that harness into explicit controls.

• model.effort on the agent's model config, set when creating the agent, per the Platform release notes.
• initial_events on POST /v1/sessions, with up to 50 user.message and user.define_outcome events, per the sessions docs.
• Up to 500 skills per session, counted across every agent in the session, per the skills docs.
• Webhooks for four environment.* lifecycle events and three memory_store.* lifecycle events, per the webhooks docs.
• Per-thread event deltas on GET /v1/sessions/{session_id}/threads/{thread_id}/stream, per Anthropic's release notes.
• Optional version on agent updates, where a supplied mismatch returns 409 and an omitted version applies unconditionally, per the agent setup docs.

Christmas came early for people building agent teams: most of the release is about removing invisible coordination tax.

The new effort setting lives inside the agent's model object, not as a session-level override. Anthropic's agent setup docs say the field accepts low, medium, high, xhigh, or max, depending on model support.

The same docs add one important constraint: an effort level inside a per-session model override is not applied, so effort is set on the agent itself. That makes it a role-level cost lever in a multi-agent roster.

Anthropic's cookbook notebook shows the pattern in miniature:

• Standards researcher: web search enabled, high effort.
• Lesson writer: web search disabled, default effort.
• Coordinator: delegates the research first, then sends the findings to the writer.

The notebook frames effort as the team's main cost lever: buy depth for the roles that need judgment, keep formatter or triage roles cheap.

The old flow created a session, then sent events to start work. The new flow lets a single create call include initial_events, processed in order.

The sessions docs put three hard edges on the feature:

• Supported event types: user.message and user.define_outcome.
• Maximum: 50 initial events.
• Non-empty list: the session is created directly in running status.

The docs also note that initial_events are not echoed on the create response. To inspect the seeded message, the examples list the session's events after creation.

Managed Agents skills are reusable, filesystem-based resources for domain-specific workflows, context, and best practices. Anthropic's skills docs split them into two types:

• Pre-built Anthropic skills: pptx, xlsx, docx, and pdf.
• Custom skills: uploaded workspace skills built around a SKILL.md file.

Each skill attachment has three fields:

• type: anthropic or custom.
• skill_id: short name for Anthropic skills, skill_* ID for custom skills.
• version: optional, defaults to latest.

The 500 limit is session-wide, counted across every agent in the session. The docs add the practical caveat: mounting more skills increases sandbox startup time.

The biggest debugging improvement is per-thread streaming for sub-agents. Anthropic's event streaming docs say event deltas render response text incrementally as a preview, while the buffered agent.message remains the authoritative record.

The cookbook's curriculum-team notebook names the old gap plainly: the session-level stream previews the primary thread, but a sub-agent's output only appeared after the whole turn was buffered. Per-thread delta streaming lets each sub-agent preview its own text live.

The notebook combines four July 22 features in one workflow:

• Per-thread delta streaming for child agent output.
• initial_events to start the session in the create call.
• model.effort per agent, used as a per-role budget lever.
• Optional version on agent update.

In a creative production workflow, that maps cleanly to a producer agent, a research agent, a draft agent, and a QA agent, each with different tools and effort.

Managed Agents already had real-time SSE streams for session work. The new webhook coverage is for lifecycle events that product teams usually poll.

The webhooks docs list the new environment events:

• environment.created
• environment.updated
• environment.archived
• environment.deleted

They also list the new memory store events:

• memory_store.created
• memory_store.archived
• memory_store.deleted

Webhook payloads carry the event type and ID, not the full object. The docs say handlers fetch the object with a GET call, deduplicate retries by the top-level event.id, and avoid relying on delivery order.

The creator-side conversation around Claude agents quickly moved from API knobs to recorded workflows. shannholmberg described a marketer teaching Claude a skill by recording the screen step by step, then using a second workstation for automations that browser agents or Playwright do not handle cleanly.

The concrete examples were not coding tasks:

• A LinkedIn interaction agent running through repetitive engagement work.
• An agent walking through competitors' product onboarding funnels and recording them.

ozansihay described the same pattern in Cowork: start a screen recording, perform the recurring task, narrate the reason for each step through the microphone, and let Claude convert the demonstration into a reusable skill. He compared it to Codex's Record & Replay feature, then singled out Claude's voice narration as the added layer. ozansihay's Cowork note

Claude Tag also picked up a governance-side control. noahzweben said admins can now attach extra auto-mode guidance when the default permission-checker classifier blocks actions they expect Claude Tag to take.

The screenshot shows the rule format: a plain sentence describing a sanctioned workflow, such as deploying to a staging cluster from a session in a channel. It also shows the limits: up to 50 rules, 1,024 characters each.