Managed Agents and Platform Tooling

In Claude Managed Agents (beta) Anthropic both runs the agent loop and provides a sandbox container for every session. You do not need to write the loop, state files or a scheduler.

20.1 Core objects

Object	What it is	How often to create
Environment	Container config: networking (limited/unrestricted), packages, MCP access	Once, reuse
Agent	Stored, versioned config: model, system prompt, tools, skills, MCP servers	Once, store the ID
Session	One run; its own workspace container; events stream	New for each task
Vault	Credentials (API keys, MCP OAuth) — substituted at egress, never visible in the sandbox	Once
Deployment	Run an agent on a cron schedule ("report every night")	As needed
Memory store	Memory that persists across sessions	As needed

Managed Agents flow

1.Environment: environments.create
2.Agent: agents.create (once)
3.Session: sessions.create (every run)
4.Events: send user.message, read the SSE stream
5.Result: files, messages, outcome

Create the agent once; never run agents.create in the request path

env = client.beta.environments.create(name="dev-env", config={
    "type": "cloud",
    "networking": {"type": "limited", "allow_package_managers": True, "allow_mcp_servers": True},
})

agent = client.beta.agents.create(          # create once, save the ID
    name="Code Reviewer", model="claude-opus-5-5",
    system="You are a strict code reviewer.",
    tools=[{"type": "agent_toolset_20260401",
            "default_config": {"enabled": True, "permission_policy": {"type": "auto"}},
            "configs": [{"name": "web_search", "enabled": False},
                        {"name": "web_fetch", "enabled": False}]}],
)

session = client.beta.sessions.create(
    agent={"type": "agent", "id": agent.id, "version": agent.version},
    environment_id=env.id,
)
with client.beta.sessions.events.stream(session_id=session.id) as stream:   # open the stream first
    client.beta.sessions.events.send(session_id=session.id, events=[
        {"type": "user.message", "content": [{"type": "text", "text": "review the auth module"}]}])
    for event in stream:
        if event.type == "agent.message":
            for b in event.content:
                if b.type == "text": print(b.text, end="")

20.2 Advanced Managed Agents features

Feature	What it does
Permission policies	always_allow, always_ask, auto (the server stops risky calls / asks for approval)
Outcomes	user.define_outcome + rubric — a separate grader agent keeps it iterating until the rubric passes
Multi-agent / workflows	A lead agent splits the work into phases across several agents (cheaper worker model)
Scheduled deployments	Autonomous sessions on cron; pause/unpause/archive
Webhooks	Notify your server on session events
Session budgets	A hard dollar cap for one session
File mounts / GitHub	Work by uploading files or mounting a repo
Self-hosted sandboxes	Tool execution on your own infra
Console session viewer	See each session's trace with no setup

20.3 ant CLI

Anthropic's command-line tool. Rule: the CLI is for the control plane (defining agents, environments), the SDK is for the data plane (running sessions).

ant auth login                 # browser OAuth; profile ~/.config/anthropic/ (Windows: %APPDATA%\Anthropic)
ant auth status                # which credential is active
ant auth login --profile admin --scope "... org:admin"     # privileged work on a separate profile

ant beta:agents list --transform '{id,name,model}' --format jsonl
ant beta:sessions:events list --session-id session_01...

ant apply                      # sync agents/*.md, environment.yaml... in the repo; IDs go in claude-lock.json

20.4 Admin, monitoring and auth

Feature	Detail
Admin API (beta)	Members, invites, workspaces, API keys, rate limit reports, service accounts — with an Admin key (sk-ant-admin...)
Usage & Cost reports	usage_report/messages, cost_report — raw HTTP; spend tracking
OpenTelemetry	The SDK emits a span for every request (C#: AddSource("com.anthropic.sdk.csharp*")); traces in Datadog/Grafana
Workload Identity Federation	CI/cloud workloads authenticate with their own identity token (JWT), without a static API key
Workspaces	Separate keys, limits, spend for teams/projects
Credential order	ANTHROPIC_API_KEY → ANTHROPIC_AUTH_TOKEN → ant profile → WIF