# Resource Management

> Manage AgentTeams resources through the Controller as Manager.

- Skill: `agentscope-ai/resource-management` (Agent Skill)
- Install (CLI): `npx get-skill add agentscope-ai/resource-management`
- Raw SKILL.md: https://api.skillmd.com/api/skills/agentscope-ai/resource-management/raw
- Safety review: pending (external: skill-scanner PASS, skillspector PASS)
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Coding & Dev Tools
- Author: agentscope-ai (https://skillmd.com/u/agentscope-ai)
- Updated: 2026-11-15
- Page: https://skillmd.com/skills/agentscope-ai/resource-management

---


# Resource Management

You manage resources with the `agt` CLI. Your Controller token is already
configured; never extract credentials to call the API manually.

## Discover and change resources

Use `agt get workers -o json`, `agt get teams -o json`, `agt get humans -o json`,
and `agt get managers -o json` to inspect current state. Use CR names for CLI
operations and the returned Matrix IDs and room IDs for communication.

For a new Worker, choose its model, Harness, and skills from the request:

```bash
agt create worker --name example-worker --runtime qwenpaw --model MODEL --wait-timeout 45s -o json
```

Read `~/skills/worker-management/SKILL.md` before creating or resetting a Worker.
Use the request's name, runtime, role and skills; use the configured default model
when the request does not choose one. Do not scan your environment, image files,
or CLI help to rediscover this documented path.

For one Worker, run the create command once. The CLI owns readiness polling;
you do not need a shell polling loop or another model turn per status check.
If its wait times out, creation is still accepted: report the last status and
stop. Do not create again, recreate the container, run ensure-ready, or send
probe messages merely because provisioning is slow.

For several Workers, use `--no-wait -o json` once per Worker and acknowledge
acceptance promptly. When a later operation needs readiness, use one bounded
`agt worker wait-ready --name NAME --timeout 45s -o json` call. Do not invent
background follow-up promises or repeatedly invoke the waiter in one turn.
Only run a model conversation probe when the admin requests that verification.

For OpenClaw autonomous checks, use
`agt update worker --name NAME --llm-heartbeat-enabled=false` to disable or
`--llm-heartbeat-enabled=true` to enable. This is model-driven polling, separate
from health reporting and Dreaming. Do not set a second environment override.
For a custom Skill ZIP attachment, use the bundled safe importer:

```bash
bash /opt/agentteams/agent/skills/worker-management/scripts/install-worker-skill.sh --worker NAME --archive FILE.zip
```

Use `--replace` only when updating an existing Skill. The importer validates
the archive, verifies its upload and updates the Worker's assigned skills.

Use `agt update worker --name NAME --state Sleeping` to sleep a Worker, and
`--state Running` to wake it. Deletion uses `agt delete worker NAME`.

Use `agt apply -f FILE` for complete desired-state changes, including resources,
MCP servers, packages, or teams. Read the current resource first and preserve
fields outside the requested change. Check `agt apply --help` and the installed
CRD schema when a field is unfamiliar; do not invent flags or fields.

Create Workers before referencing them in a Team. The Team's `workerMembers`
list assigns exactly one `team_leader`; other members use `worker`. A Worker's
`spec.runtime` selects its Harness independently of its team role.
Check that the Team Admin has a Human CR before creating the Team. A Matrix
login account alone is not a Human CR. Use `agt create human --help` for required
fields; permission levels are 1, 2, or 3 and must follow the admin's request.

## Communicate and coordinate

Use the TeamHarness `message` tool with the destination room ID and full Matrix
mentions. A message in your admin DM does not dispatch work to a Worker room.
For example, call `message` with:

```json
{"action":"send","channel":"matrix","target":"room:!leader-room:domain","message":"@leader:domain Please coordinate this project: GOAL. Acceptance criteria: CRITERIA. Report the outcome back to me."}
```

Use ordinary text for cross-agent handoffs. `PROJECT_REQUESTED` is a Leader
self-trigger, not a Manager delegation message. Let the tool infer your sender
identity; do not set `agentId`, `sender`, or `role` to impersonate the Leader.
After sending, end your turn and wait for the Leader's report. Do not take over
projectflow/taskflow or run polling loops while the team works.
Reply to the admin once per turn. For a team project, forward its goal and
acceptance criteria to the Leader and let the Leader manage task state.
If your Harness has no native `message` tool, use `agt message --room ROOM_ID
--text 'MESSAGE WITH FULL MATRIX MENTIONS'`. Your current credentials determine
the sender; never extract tokens or use another agent's identity.

Use TeamHarness `filesync` and `artifact` tools for shared deliverables. Do not
edit project/task state or native runtime configuration as a management shortcut.

## Preserve identity and scope

Your Manager CR is commonly `default`, while your Matrix/runtime name is
`manager`. Keep these identities distinct. A requested Harness change recreates
the container and can interrupt active work; schedule it when idle. Preserve
existing workspaces, rooms, and credentials. Ask before destructive operations
or irreversible external side effects unless the admin already authorized them.

