docs(part5): add a complete multi-agent worked example (#13)#19
Conversation
Add a Chief of Staff + workers example to Part 5: file layout (SOUL/AGENTS/TOOLS per agent), an agents.list openclaw.json config, Task Brain semantic-approval scoping per agent, the orchestrator AGENTS.md delegation rulebook, the runtime spawn flow, and a token/latency optimization table. Grounded in templates/openclaw.example.json shapes, with a 'verify keys against your version' caveat consistent with #8.
🤖 Devin AI EngineerI'll be helping with this pull request! Here's what you should know: ✅ I will automatically:
Note: I can only respond to comments from users who have write access to this repository. ⚙️ Control Options:
|
![devin-ai-integration[bot]](https://avatars.githubusercontent.com/in/811515?s=60&v=4){kind=small}
| { | ||
| "taskBrain": { | ||
| "approvals": { | ||
| "defaults": { "read-only.*": "allow", "execution.*": "ask", "write.network": "ask", "control-plane.*": "deny" } |
There was a problem hiding this comment.
🟡 taskBrain approval defaults missing critical categories present in the referenced template
The approval defaults shown in this config excerpt omit three categories that are present in templates/openclaw.example.json:191-198: "execution.sandbox.*": "allow", "write.fs.workspace": "allow", and "write.fs.outside-workspace": "deny". This matters because the prose at line 910 then tells users the coding-worker needs write.fs.workspace = allow, but that category doesn't appear anywhere in the shown config — neither in the defaults block nor as a per-agent override. A user copying this as their starting config would get a policy where sandboxed execution prompts for approval (falling through to execution.*: ask) and workspace writes have no explicit rule at all. The text at line 847 claims these configs use "the config shapes from templates/openclaw.example.json", making the omission a concrete inconsistency rather than an intentional simplification.
| "defaults": { "read-only.*": "allow", "execution.*": "ask", "write.network": "ask", "control-plane.*": "deny" } | |
| "defaults": { "read-only.*": "allow", "execution.sandbox.*": "allow", "execution.*": "ask", "write.fs.workspace": "allow", "write.fs.outside-workspace": "deny", "write.network": "ask", "control-plane.*": "deny" } |
Was this helpful? React with 👍 or 👎 to provide feedback.
There was a problem hiding this comment.
Fixed in 209e4b6 — restored the full defaults block (execution.sandbox.*, write.fs.workspace, write.fs.outside-workspace) to match templates/openclaw.example.json:191-198, so the config is copy-paste-safe and the prose that references write.fs.workspace now lines up with what's shown. Thanks.
The abbreviated taskBrain.approvals.defaults omitted execution.sandbox.*, write.fs.workspace, and write.fs.outside-workspace that the surrounding prose references and that templates/openclaw.example.json ships (per Devin Review on #19).
2 participants
Summary
Closes #13. Part 5 explained the orchestration patterns (sub-agents-as-GC, Anthropic's 5 coordination patterns, CEO/COO/Worker) but never showed a single end-to-end multi-agent system wired together. This adds a Chief of Staff + two workers worked example to Part 5:
SOUL.md/AGENTS.md/TOOLS.mdper agent (the three files that survive the 2026.5.22 worker-bootstrap narrowing).openclaw.jsonconfig —agents.defaults+agents.list[]with a frontier orchestrator and cheaper/local workers (one in lean mode).tools.toolsBySender, so workers are enforced into their lane rather than prompt-trusted.AGENTS.md— the actual delegation rulebook (when to spawn, self-contained work orders, Generator-Verifier verification).openclaw tasks list.cache-ttlpruning, cheap workers, lean mode, cheap compaction model).Everything is built from the config shapes already in
templates/openclaw.example.json— no new tools or claims invented.Type of change
Review checklist
markdownlint-cli2andlychee --offlinelocally (CI will also run these)Notes / flagged items
taskBrain.approvals.defaultsblock verbatim from the template, then describes the per-agent intent in prose and explicitly tells the reader to confirm the exact per-agent nesting against their version (consistent with the Part 24: 'openclaw flows' CLI and semantic approval categories don't exist in 2026.4.15 / 2026.4.19-beta.2 #8 version-accuracy caveat, which this section links to). I deliberately avoided inventing a preciseagents.list[].taskBrain.approvalskey shape.Link to Devin session: https://app.devin.ai/sessions/21e86e63e6ed4a7cb88868ac2243d9ef
Requested by: @OnlyTerp