agi-farm

Builds a complete multi-agent AI team on OpenClaw. One wizard, full team.

Commands

/agi-farm setup

Ask one question at a time. Do not proceed until confirmed.

Step 1 — Team name

"What should we call your team? (e.g. NovaCorp, TradingDesk — default: MyTeam)"

Store as TEAM_NAME.

Step 2 — Orchestrator name

"What's your orchestrator's name? (default: Cooper)"

Store as ORCHESTRATOR_NAME.

Step 3 — Team size

"How many agents? 3 — Minimal: Orchestrator + Researcher + Builder 5 — Standard: adds QA + Content 11 — Full stack: complete AGI system (recommended)"

Store as PRESET.

Step 3.5 — Domain

"What domain? software / trading / research / general (default) / custom"

If custom: ask for one-phrase description. Store as DOMAIN.

Step 3.6 — Custom agents (PRESET 3 or 5 only)

"Add a custom agent? (yes/no, default: no)"

If yes, collect per agent: id, name, emoji, role, goal. Max 3 custom agents. Append to roster in Step 7 with "template": "generic".

Step 4 — Frameworks

"Collaboration frameworks? autogen / crewai / langgraph / all / none"

Store as FRAMEWORKS list. all → ["autogen", "crewai", "langgraph"].

Step 5 — GitHub

"Create a GitHub repo for the bundle? yes / no"

Store as CREATE_GITHUB.

Step 6 — Confirm

Show summary, ask "Shall I proceed? (yes/no)". If no → restart Step 1.

Step 7 — Write team.json

mkdir -p ~/.openclaw/workspace/agi-farm-bundle/
openclaw agents list --json   # use output to assign appropriate models per role

Use the openclaw agents list output to assign each agent a model appropriate for its role. Write resolved model strings directly into the "model" fields.

Model selection cheat sheet (based on openclaw agents list --json output):

Tip: assign opus or sonnet to roles that make decisions; use flash/glm-4.7-flash for high-frequency reviewers to manage cost.

3-agent roster:

{"team_name":"\x3CTEAM_NAME>","orchestrator_name":"\x3CORCHESTRATOR_NAME>","preset":"3",
 "domain":"\x3CDOMAIN>","frameworks":\x3CFRAMEWORKS_JSON>,"created_at":"\x3CISO_TIMESTAMP>",
 "agents":[
   {"id":"main",       "name":"\x3CORCHESTRATOR_NAME>","emoji":"🦅","role":"Orchestrator",           "goal":"Orchestrate the team, delegate tasks, synthesize results",             "model":"\x3CMODEL>","workspace":"."},
   {"id":"researcher", "name":"Sage",               "emoji":"🔮","role":"Researcher",             "goal":"Research deeply and surface the insights that matter most",            "model":"\x3CMODEL>","workspace":"researcher"},
   {"id":"builder",    "name":"Forge",              "emoji":"⚒️","role":"Builder",                "goal":"Implement solutions cleanly and efficiently",                          "model":"\x3CMODEL>","workspace":"builder"}
 ]}

5-agent: add to 3-agent roster:

{"id":"qa",     "name":"Vigil", "emoji":"🛡️","role":"QA Engineer",       "goal":"Ensure every output meets quality standards","model":"\x3CMODEL>","workspace":"qa"},
{"id":"content","name":"Anchor","emoji":"⚓", "role":"Content Specialist","goal":"Craft clear content that communicates complex ideas simply","model":"\x3CMODEL>","workspace":"content"}

11-agent roster:

[
  {"id":"main",  "name":"\x3CORCHESTRATOR_NAME>","emoji":"🦅","role":"Orchestrator",            "goal":"Orchestrate specialists, delegate tasks, synthesize results",                          "model":"\x3CMODEL>","workspace":"."},
  {"id":"sage",  "name":"Sage",              "emoji":"🔮","role":"Solution Architect",       "goal":"Design robust, scalable architectures",                                               "model":"\x3CMODEL>","workspace":"solution-architect"},
  {"id":"forge", "name":"Forge",             "emoji":"⚒️","role":"Implementation Engineer", "goal":"Implement clean, well-tested code efficiently",                                        "model":"\x3CMODEL>","workspace":"implementation-engineer"},
  {"id":"pixel", "name":"Pixel",             "emoji":"🐛","role":"Debugger",                "goal":"Find the true root cause of any bug or failure",                                       "model":"\x3CMODEL>","workspace":"debugger"},
  {"id":"vista", "name":"Vista",             "emoji":"🔭","role":"Business Analyst",        "goal":"Research deeply and surface the insights that matter most",                            "model":"\x3CMODEL>","workspace":"business-analyst"},
  {"id":"cipher","name":"Cipher",            "emoji":"🔊","role":"Knowledge Curator",       "goal":"Curate and surface knowledge so the team never forgets",                              "model":"\x3CMODEL>","workspace":"knowledge-curator"},
  {"id":"vigil", "name":"Vigil",             "emoji":"🛡️","role":"QA Engineer",             "goal":"Ensure every output meets quality standards",                                         "model":"\x3CMODEL>","workspace":"quality-assurance"},
  {"id":"anchor","name":"Anchor",            "emoji":"⚓", "role":"Content Specialist",     "goal":"Craft clear content that communicates complex ideas simply",                           "model":"\x3CMODEL>","workspace":"content-specialist"},
  {"id":"lens",  "name":"Lens",              "emoji":"📡","role":"Multimodal Specialist",   "goal":"Extract meaning from images, documents, and multimodal inputs",                       "model":"\x3CMODEL>","workspace":"multimodal-specialist"},
  {"id":"evolve","name":"Evolve",            "emoji":"🔄","role":"Process Improvement Lead","goal":"Make the team better systematically through continuous improvement",                   "model":"\x3CMODEL>","workspace":"process-improvement"},
  {"id":"nova",  "name":"Nova",              "emoji":"🧪","role":"R&D Lead",                "goal":"Turn hypotheses into proven capabilities through structured experimentation",          "model":"\x3CMODEL>","workspace":"r-and-d"}
]

Step 8 — Generate workspace files

python3 ~/.openclaw/skills/agi-farm/generate.py \
  --team-json ~/.openclaw/workspace/agi-farm-bundle/team.json \
  --output ~/.openclaw/workspace/ \
  --all-agents --shared --bundle

Step 9 — Create OpenClaw agents

For each agent except main (skip if already exists):

openclaw agents add \
  --agent \x3Cid> --name "\x3Cname>" --emoji "\x3Cemoji>" \
  --model "\x3Cmodel>" \
  --workspace "~/.openclaw/workspace/agents-workspaces/\x3Cworkspace>"

Use agent["model"] from team.json directly.

Step 10 — Register cron jobs

python3 ~/.openclaw/skills/agi-farm/scripts/register-crons.py \
  --team-json ~/.openclaw/workspace/agi-farm-bundle/team.json

Timezone is read automatically from OpenClaw config. Skips any cron that already exists.

Step 11 — Install frameworks

For each framework in FRAMEWORKS:

if [ ! -d ~/.openclaw/skills/\x3Cfw>-collab ]; then
  TMP=$(mktemp -d)
  git clone --depth 1 --filter=blob:none --sparse \
    https://github.com/oabdelmaksoud/openclaw-skills.git "$TMP"
  cd "$TMP" && git sparse-checkout set \x3Cfw>-collab
  cp -r \x3Cfw>-collab ~/.openclaw/skills/ && rm -rf "$TMP"
fi
python3 ~/.openclaw/skills/\x3Cfw>-collab/build_agents.py --force 2>/dev/null || true

Step 12 — GitHub (if chosen)

cd ~/.openclaw/workspace/agi-farm-bundle
git init -b main && git add . && git commit -m "feat: \x3CTEAM_NAME> AGI farm"
gh repo create agi-farm-\x3CTEAM_NAME_LOWER> --public --source . --remote origin --push

Step 13 — Commit workspace

cd ~/.openclaw/workspace
git add -A && git commit -m "feat: \x3CTEAM_NAME> AGI team — agi-farm setup complete"

Step 14 — Initialize registries + health check

# Write TASKS.json and AGENT_STATUS.json
python3 - \x3C\x3C 'EOF'
import json
from pathlib import Path
ws   = Path.home() / ".openclaw/workspace"
team = json.loads((ws / "agi-farm-bundle/team.json").read_text())
(ws / "TASKS.json").write_text("[]")
(ws / "AGENT_STATUS.json").write_text(json.dumps(
    {a["id"]: {"status": "available", "name": a["name"]} for a in team["agents"]}, indent=2))
print("✅ registries written")
EOF

# Health check
AGENTS=$(openclaw agents list --json 2>/dev/null | python3 -c "import json,sys; print(len(json.load(sys.stdin)))" || echo 0)
CRONS=$(openclaw cron list 2>/dev/null | grep -c "\x3CTEAM_NAME_LOWER>" || echo 0)
[ -d ~/.openclaw/workspace/comms/inboxes ] && echo "✅ comms OK" || echo "❌ comms missing"
[ -f ~/.openclaw/workspace/TASKS.json ]    && echo "✅ TASKS.json OK" || echo "❌ TASKS.json missing"
echo "✅ Agents: $AGENTS | Crons: $CRONS"

Step 15 — Done

✅ \x3CTEAM_NAME> AGI team is live!
Agents   : \x3CPRESET> (\x3CAGENT_NAMES_LIST>)
Workspace: ~/.openclaw/workspace/
Bundle   : ~/.openclaw/workspace/agi-farm-bundle/
GitHub   : \x3CURL if created>

Next: talk to \x3CORCHESTRATOR_NAME> · /agi-farm status · /agi-farm dashboard

/agi-farm status

openclaw agents list --json | python3 -c "
import json,sys
for a in json.load(sys.stdin):
    print(f'  {a.get(\"identityEmoji\",\"🤖\")} {a.get(\"identityName\",a[\"id\"])}: {a.get(\"model\",\"?\")}')
"
python3 -c "
import json
from pathlib import Path
ws = Path.home() / '.openclaw/workspace'
tasks = json.loads((ws/'TASKS.json').read_text()) if (ws/'TASKS.json').exists() else []
t = [t for t in tasks if isinstance(t,dict)]
print(f'  Tasks: {len(t)} total · {sum(1 for x in t if x.get(\"status\")==\"pending\")} pending · {sum(1 for x in t if x.get(\"status\")==\"needs_human_decision\")} HITL')
"
openclaw cron list 2>/dev/null | head -15

/agi-farm rebuild

python3 ~/.openclaw/skills/agi-farm/generate.py \
  --team-json ~/.openclaw/workspace/agi-farm-bundle/team.json \
  --output ~/.openclaw/workspace/ \
  --all-agents --shared --no-overwrite

--no-overwrite skips files that already exist, preserving manual edits. Add --force (remove --no-overwrite) to overwrite everything.

/agi-farm export

cd ~/.openclaw/workspace/agi-farm-bundle
git add -A
git commit -m "export: $(date +%Y-%m-%d)" 2>/dev/null || echo "Nothing to commit"
git push 2>/dev/null || echo "No remote — run /agi-farm setup first"

/agi-farm dashboard

React + SSE ops room. File-watcher pushes live data to the browser in ~350ms on any workspace .json or .md change. Runs as a persistent macOS LaunchAgent — always on, auto-restarts on crash.

Architecture

dashboard.py          ← Python HTTP server (SSE + static)
  ├── WorkspaceWatcher  watchdog file-watcher, 250ms debounce
  ├── SlowDataCache     background thread — caches `openclaw agents list`
  │                     and `openclaw cron list` every 30s (each takes ~1-2s)
  ├── Broadcaster       thread-safe SSE fan-out to all connected clients
  └── /api/stream       SSE endpoint — pushes full snapshot on every file change

dashboard-react/      ← Vite + React 18 + Recharts frontend
  dist/               ← production build (served by dashboard.py)
  src/
    hooks/useDashboard.js   SSE hook — auto-reconnects on disconnect
    components/
      Header.jsx            live badge, stats, clock
      Nav.jsx               tab switcher
      tabs/
        Overview.jsx        stats, budget bar, SLA alerts, agent grid, broadcast preview
        Agents.jsx          full agent cards — model, inbox, quality, credibility, cache age
        Tasks.jsx           filterable table, expandable rows, ticking deadlines, pagination
        Velocity.jsx        7-day charts (Recharts), quality trend, task-type donut
        Budget.jsx          period bars, threshold markers, per-agent/model breakdown
        OKRs.jsx            objectives + KRs with progress bars
        RD.jsx              experiments, backlog, benchmarks
        Broadcast.jsx       terminal log, color-coded CRITICAL/BLOCKED/HITL

Data sources (all real-time from workspace files)

LaunchAgent (always-on)

The dashboard is registered as ai.coopercorp.dashboard and starts automatically at login.

# Status
launchctl list | grep coopercorp
curl -s http://localhost:8080/api/data | python3 -m json.tool | head -5

# Restart
launchctl stop  ai.coopercorp.dashboard
launchctl start ai.coopercorp.dashboard

# Logs
tail -f /tmp/coopercorp-dashboard.log
tail -f /tmp/coopercorp-dashboard.err

# Disable / re-enable
launchctl unload ~/Library/LaunchAgents/ai.coopercorp.dashboard.plist
launchctl load   ~/Library/LaunchAgents/ai.coopercorp.dashboard.plist

URL: http://localhost:8080

Rebuild React frontend

cd ~/.openclaw/skills/agi-farm/dashboard-react
npm install        # first time only
npm run build      # outputs to dist/ — dashboard.py serves automatically

Full reference: references/dashboard.md

/agi-farm dispatch

# Dry-run (preview only)
python3 ~/.openclaw/skills/agi-farm/scripts/auto-dispatch.py

# Execute
python3 ~/.openclaw/skills/agi-farm/scripts/auto-dispatch.py --execute

Fires agent sessions for pending tasks, handles HITL notifications, stale task resets, rate-limit backoff, and dependency checking. Cron (every 1 min):

* * * * * python3 ~/.openclaw/skills/agi-farm/scripts/auto-dispatch.py --execute \
            >> ~/.openclaw/workspace/logs/auto-dispatch.log 2>&1

Troubleshooting

Setup issues

Runtime issues

Recovery

# Re-run setup without overwriting existing files
python3 ~/.openclaw/skills/agi-farm/generate.py \
  --team-json ~/.openclaw/workspace/agi-farm-bundle/team.json \
  --output ~/.openclaw/workspace/ \
  --all-agents --shared --no-overwrite

# Force full regeneration (overwrites everything)
python3 ~/.openclaw/skills/agi-farm/generate.py \
  --team-json ~/.openclaw/workspace/agi-farm-bundle/team.json \
  --output ~/.openclaw/workspace/ \
  --all-agents --shared --bundle --force