Claude Managed Agents

Use this skill to operate Anthropic Claude Managed Agents safely and cleanly from this machine.

This skill is SDK-first through the bundled Python helper, with ant CLI documented as a secondary operator lane. Raw REST is only for debugging or when the SDK is unavailable.

What this skill handles

• agents
• create
• update
• retrieve
• list
• list versions
• archive
• delete
• environments
• create
• update
• retrieve
• list
• archive
• delete
• sessions
• create
• retrieve
• list
• archive
• delete
• session events
• send user messages
• interrupt and redirect
• list history
• stream live SSE events
• send tool confirmations
• send custom tool results
• files
• upload local files to the Files API
• list, download, and delete files
• return file IDs for mounting into sessions
• session resources
• add resources to running sessions
• list mounted resources
• delete mounted resources by resource ID
• configuration domains
• built-in toolset controls
• MCP servers
• skills
• packages
• networking
• vault IDs
• mounted resources
• diagnostics
• local preflight via doctor
• optional live read-only connectivity checks via doctor --live

Required environment

Set:

• ANTHROPIC_API_KEY

Optional:

• ANTHROPIC_API_BASE_URL
• ANTHROPIC_MANAGED_AGENTS_BETA
• ANTHROPIC_TIMEOUT_SECONDS

The managed-agents beta header is required. The helper uses managed-agents-2026-04-01 by default.

Script

Use:

• scripts/managed_agents.py

Run it with Python 3:

python3 ~/.openclaw/skills/claude-managed-agents/scripts/managed_agents.py --help

Run a preflight before live work when the lane feels sketchy:

python3 ~/.openclaw/skills/claude-managed-agents/scripts/managed_agents.py \
  doctor \
  --allowed-host api.example.com

Operating model

Prefer this order:

1. Python helper CLI
2. documented ant equivalent when the user wants a direct CLI path
3. raw REST only when troubleshooting edge cases

The helper supports:

• --backend auto (default)
• --backend sdk
• --backend http

Use sdk when the Anthropic Python SDK is installed. Use http if the SDK is missing or behaving oddly.

Safe workflow

For new setups

1. confirm ANTHROPIC_API_KEY exists
2. create or inspect the target agent
3. create or inspect the target environment
4. create the session
5. send a user message event
6. stream or list events to monitor progress

For mid-run steering

1. list or stream session events
2. if the agent is going the wrong direction, send:
• user.interrupt
• followed by a new user.message
3. if the session is waiting on approval, send user.tool_confirmation
4. if the session is waiting on a custom tool result, send user.custom_tool_result

High-value guardrails

• Do not guess event IDs, tool use IDs, or custom tool use IDs.
• For approval flows, read recent events first and use the exact pending tool ID.
• For custom tools, return only the result the tool actually produced. Do not fabricate success.
• Prefer limited networking with explicit hosts for production-oriented environments. Use bare hostnames like api.example.com, not full URLs.
• Do not delete environments, sessions, or agents casually. Archive first unless the user clearly wants hard deletion.
• Archiving is usually the safer lifecycle move than deletion, but disposable smoke fixtures may need delete for full cleanup.

Recommended command patterns

Create an agent with the full built-in toolset

python3 ~/.openclaw/skills/claude-managed-agents/scripts/managed_agents.py \
  agent create \
  --name "Coding Assistant" \
  --model claude-sonnet-4-6 \
  --system "You are a helpful coding agent." \
  --agent-toolset

Run doctor before a live session

python3 ~/.openclaw/skills/claude-managed-agents/scripts/managed_agents.py \
  doctor \
  --live \
  --allowed-host api.example.com

Create an environment with limited networking and pip packages

python3 ~/.openclaw/skills/claude-managed-agents/scripts/managed_agents.py \
  environment create \
  --name "python-dev" \
  --network limited \
  --allowed-host api.example.com \
  --allow-package-managers \
  --pip pandas==2.2.0 \
  --pip numpy==2.1.0

Upload a file and get a file ID

python3 ~/.openclaw/skills/claude-managed-agents/scripts/managed_agents.py \
  file upload \
  --file-path ./data.csv \
  --only-id

Create a session with an uploaded file mounted in the container

FILE_ID=$(python3 ~/.openclaw/skills/claude-managed-agents/scripts/managed_agents.py \
  file upload \
  --file-path ./data.csv \
  --only-id)

python3 ~/.openclaw/skills/claude-managed-agents/scripts/managed_agents.py \
  session create \
  --agent-id agent_123 \
  --environment-id env_123 \
  --title "Repo analysis" \
  --resource-json "{\"type\":\"file\",\"file_id\":\"${FILE_ID}\",\"mount_path\":\"/workspace/data.csv\"}"

Add another file to a running session

python3 ~/.openclaw/skills/claude-managed-agents/scripts/managed_agents.py \
  session resource add \
  --session-id sess_123 \
  --file-id file_abc123 \
  --mount-path /workspace/config.json

Download a session-scoped file artifact

python3 ~/.openclaw/skills/claude-managed-agents/scripts/managed_agents.py \
  file download \
  --file-id file_abc123 \
  --output ./artifact.txt

Use this mainly for generated artifacts. Uploaded source files often come back with downloadable: false.

Send a user message to the session

python3 ~/.openclaw/skills/claude-managed-agents/scripts/managed_agents.py \
  session send \
  --session-id sess_123 \
  --message "Summarize the repository and propose the next refactor."

Interrupt and redirect the session

python3 ~/.openclaw/skills/claude-managed-agents/scripts/managed_agents.py \
  session send \
  --session-id sess_123 \
  --interrupt \
  --message "Stop the broad audit and focus on the auth bug in line 42."

Stream events until idle

python3 ~/.openclaw/skills/claude-managed-agents/scripts/managed_agents.py \
  session stream \
  --session-id sess_123 \
  --until-idle

Approve a pending tool call

python3 ~/.openclaw/skills/claude-managed-agents/scripts/managed_agents.py \
  session send \
  --session-id sess_123 \
  --confirm-tool-use-id tool_evt_123 \
  --confirm-result allow

Delete an agent when the user explicitly wants permanent cleanup

python3 ~/.openclaw/skills/claude-managed-agents/scripts/managed_agents.py \
  agent delete \
  --agent-id agent_123

Return a custom tool result

python3 ~/.openclaw/skills/claude-managed-agents/scripts/managed_agents.py \
  session send \
  --session-id sess_123 \
  --custom-tool-use-id custom_evt_123 \
  --custom-tool-text '{"temperature_f":72,"condition":"sunny"}'

When to use ant CLI instead

Use ant when the user explicitly wants Anthropic's native CLI experience, copy-pasteable operator commands, or quick manual inspection.

Examples:

ant beta:agents create --name "Coding Assistant" --model '{id: claude-sonnet-4-6}' --tool '{type: agent_toolset_20260401}'
ant beta:environments list
ant beta:sessions retrieve --session-id "$SESSION_ID"
ant beta:sessions:events send --session-id "$SESSION_ID"
ant beta:sessions stream --session-id "$SESSION_ID"

References in this skill

Read these when needed:

• references/quickstart.md for the end-to-end happy path
• references/lifecycle-recipes.md for lifecycle operations and payload patterns
• references/files-api.md for upload, list, download, delete, and session resource workflows
• references/event-model.md for streaming, interruptions, approvals, and custom tools
• references/ant-cli-recipes.md for direct ant commands mirroring the helper
• references/known-gaps.md for dependency assumptions and operational caveats

Output contract

Default answer shape:

• short lead sentence with the answer or current state
• 2-6 bullets with the important IDs, statuses, or lifecycle facts
• exact next command when a follow-up step is likely

If the user asks for raw JSON, return the raw JSON instead.