Claude Managed Agents

The headless deployment surface for claude-financial-services — Anthropic's Managed Agents API (/v1/agents), the counterpart to claude-cowork in the "two ways from one source" model.

In this wiki

• Each template under managed-agent-cookbooks/ references the same system prompt and skills as its Cowork plugin counterpart, plus an agent.yaml, depth-1 leaf-worker subagents, and steering-event examples.
• A deploy script resolves file references, uploads skills, creates the subagents, and POSTs an orchestrator to /v1/agents. A reference event loop routes handoff_request events between agents via the firm's own orchestration layer.
• Subagent delegation (callable_agents) is a Research Preview capability.
• Built on model-context-protocol connectors for data access.

Its leaf-worker subagents + handoff_request loop are an instance of agent-orchestration.

Terminal client — the ant CLI

Beyond the FSI deploy-script pattern, Anthropic now ships a first-party CLI, ant (ant-cli), that drives this /v1/agents surface directly: create → deploy to Anthropic's cloud → send tasks → pull execution traces, "without creating any wrapper code." It turns Managed Agents into a terminal primitive — the developer-facing edge of anthropic's model-provider → full-agent-platform shift.

Related

claude-financial-services · claude-cowork · ant-cli · claude-dreams · model-context-protocol · agent-orchestration · agent-memory · agent-scope-isolation · durable-agents · anthropic

Sourced from the financial-services README + the ant-cli hands-on (the CLI client).

Re-verified 2026-08-04 (staleness queue)

url: added (platform docs). Status: public beta, behind the managed-agents-2026-04-01 beta header, with memory-store endpoints on a separate agent-memory-2026-07-22 header (corrected 2026-09-22: the docs now say "session and memory-store calls need only managed-agents-2026-04-01"; only dreams keep a second header — claude-dreams).

Re-verified 2026-09-22

Still beta, enabled by default for all API accounts. Three additions since the August read. An environment can now be a self-hosted sandbox on the customer's own infrastructure, not only Anthropic's cloud sandbox. Scheduled deployments run an agent on a cron schedule. And the service is also offered on Claude Platform on AWS, with some differences. Two limits are stated plainly: MCP tunnels and dreaming sit in a narrower research preview behind a request-access form, and because sessions keep state on the server, Managed Agents is not eligible for Zero Data Retention or HIPAA BAA coverage.

Two design facts worth carrying into this spoke's agent-memory and agent-scope-isolation threads. An agent is a persisted, versioned config — model, system prompt, tools, MCP servers and skills are top-level fields on POST /v1/agents, never on a session — and every update creates a new immutable version, with sessions pinning to a version at creation. And there is no delete, only archive, which is permanent: the agent becomes read-only, no new sessions can reference it, and there is no unarchive. That is a stronger immutability guarantee than any other agent runtime in this corpus offers, and it is the opposite instinct to the mutable-markdown memory pole.

Re-verified 2026-08-22 — the overview doc (hub-routed, Telegram)

The platform overview page, read 2026-08-22, is the first time this spoke holds Anthropic's own framing of the product rather than a deploy script or a CLI hands-on. Anthropic positions it as one of two ways to build: the Messages API for "custom agent loops and fine-grained control", Managed Agents for "long-running tasks and asynchronous work." Its own words for what it sells: "Instead of building your own agent loop, tool execution, and runtime, you get a fully managed environment."

Four objects.

Config and placement are separate objects. That means the same versioned agent can run in Anthropic's sandbox or on your hardware without being redefined, which is how the docs answer "compliance or data-residency requirements." No other runtime in this spoke separates the two that cleanly — it is the agent-scope-isolation instinct applied to infrastructure rather than to principals.

Built-in tools: bash; file read/write/edit/glob/grep; web search and fetch, optionally restricted to an allowlist or blocklist of domains; MCP servers. The docs describe the harness as shipping "built-in prompt caching, compaction, and other performance optimizations" — so context-rot management is a platform feature here, not something the developer builds.

Scheduled deployments run sessions on a cron cadence with no client-side scheduler. That closes the loop with durable-agents and with the unattended-operator thread: a hosted agent that wakes itself is the same shape as mimiclaw's 30-minute heartbeat, with the vendor owning the timer.

Status and the honest caveat. Public beta behind managed-agents-2026-04-01, enabled by default for all API accounts. Two things sit behind a further research preview and a separate request form: MCP tunnels and dreaming (claude-dreams). And because sessions are "stateful by design" — conversation history, sandbox state and outputs all stored server-side — Managed Agents is not eligible for Zero Data Retention or a HIPAA BAA. That is the price of the managed pole stated plainly by the vendor, and it is the same trade cloudflare-agent-memory makes: the state that makes the service useful is the state you no longer hold. Anthropic's mitigation is deletion on demand (sessions and uploaded files, via the API); Cloudflare's is exportability. Both are policy, not architecture.

Also available on Claude Platform on AWS, with the docs flagging "some differences in feature availability and session behavior" — unspecified on this page.