agent-team-manager - SKILL.md Agent Skill

name: agent-team-manager description: Offload any non-trivial task to a background agent by default so the main agent stays free; persist named domain experts on request type: agent-skill triggers: - "/atm" - "/agent-team-manager" - "assign this task to" - "create an agent for" - "make a dedicated" - "background agent" - "agent team" - "task delegation" metadata: category: "workflow-management" tags: ["agent", "task", "delegation", "background", "team"] version: "3.0.0" author: "OpenClaw"

Agent Team Manager

Default posture: offload, don't block. If a task would occupy you (the main agent) for more than a few seconds, hand it to a background agent and stay available for the next request. Only trivial, near-instant work is handled inline.

The decision rubric — apply to EVERY task

Handle inline only if ALL of these are true:

answerable in 0–1 tool calls
finishes in ~a few seconds
no file edits, no repeated external/shell calls, not multi-step

Examples: a fact or definition, a unit conversion, a single quick file read, a one-line edit, "what time is it".

Offload to a background agent if ANY of these are true:

likely needs ≥2 tool calls or >~10 seconds
edits files, hits the network/shell repeatedly, or is multi-step/ongoing

Examples: research, code changes across files, builds/tests, data analysis, drafting documents, anything investigative or open-ended.

When unsure → offload. Bias toward keeping yourself free. The whole point is that a 30-second task must not block the user's next request.

How to offload

Decide the domain (research, code, design, writing, analytics, market, customer, or general).
Spawn it as a background task using your native background-task / subagent tool, with notify policy = completion + errors so results and failures are pushed back to the user automatically.
The skill's helper records the task. Invoke: node agent-team-manager.js task "<the task>" This spawns the background worker, logs it, and tracks domain recurrence.
Immediately tell the user you've offloaded it and are free for the next request. Do not wait for the background task inline.

Hybrid persona model

Default = ephemeral worker. One-off tasks get a fire-and-forget background worker. No persistent persona, nothing to manage.
Persist on request. When the user says "make a dedicated agent for X" (or similar), create/reuse a named domain expert: node agent-team-manager.js task --persist "<the task>"
Promote on recurrence. On the user's 3rd task in the same domain, the helper surfaces an offer to promote it to a standing named agent. Pass the offer along; only create the persistent agent if the user accepts.

Notifications

All background tasks notify on completion and errors. Completed results are delivered back to the user automatically; failures surface proactively. You never need to poll inline.

Commands

node agent-team-manager.js task "<task>"            # offload (default, ephemeral)
node agent-team-manager.js task --persist "<task>"  # offload to a persistent named agent
node agent-team-manager.js tasks                    # list background tasks + status
node agent-team-manager.js done <id>                # mark a background task completed (keeps the log + prune accurate)
node agent-team-manager.js fail <id>                # mark a background task failed
node agent-team-manager.js list                     # list persistent named agents
node agent-team-manager.js query <agent> "<q>"      # ask a persistent agent

Tracking completion

OpenClaw's background-task runtime notifies on completion and errors. When a background task finishes, mark it in the skill's log so tasks listings and the 7-day auto-prune stay accurate:

node agent-team-manager.js done <id> (completed)
node agent-team-manager.js fail <id> (failed)

State

agents.json — persistent named agents only (name, domain, status, lastUsed, taskCount) plus per-domain recurrence counts.
tasks.json — ephemeral background-task log (auto-pruned after 7 days).

Security

Background agents run in isolated sessions; no agent reads another's context.
All spawns are logged in tasks.json.