Crewai Collab

crewai-collab — Structured Multi-Agent Task Runner

Uses real CrewAI 1.9.3 orchestration with role-based agents and typed tasks. Routes all LLM calls through openclaw agent --json — no API keys required.

Use this (not autogen-collab) when:

• Tasks have a clear pipeline structure (research → design → implement)
• You need explicit expected outputs per task ("return a class diagram")
• You want a manager agent (Cooper/main) reviewing and delegating each step
• Tasks have defined acceptance criteria

Use autogen-collab instead when:

• Open-ended debate with no fixed structure
• You just need the best answer, not a specific artifact format

Prerequisites (first run only)

python3 ~/.openclaw/skills/crewai-collab/build_agents.py --force

Run once after setup or after updating any agent's SOUL.md. The runner auto-rebuilds stale configs on each invocation, but the agents directory must exist.

Process Types

Available Agents

How to Invoke

1. Generate a task ID

import uuid; task_id = str(uuid.uuid4())[:8]

2. Choose agents + process

3. Run the crew (background process)

~/.openclaw/skills/crewai-collab/.venv/bin/python3 \
  ~/.openclaw/skills/crewai-collab/crewai_runner.py \
  --process sequential \
  --agents vista,sage,forge \
  --tasks "Research|Research the problem domain and identify key constraints|Summary with 5 key findings and constraints" \
          "Design|Design a solution architecture based on the research findings|Architecture spec with component diagram" \
          "Implement|Write a Python implementation of the designed architecture|Working Python module with docstrings" \
  --task-id "<task-id>" \
  --output /Users/omarabdelmaksoud/.openclaw/workspace/comms/crewai/<task-id>/ \
  --turn-timeout 90 \
  --timeout 300

Use OpenClaw's exec tool with background=true so Cooper doesn't block.

4. Poll for completion

Every 15–30 seconds:

cat /Users/omarabdelmaksoud/.openclaw/workspace/comms/crewai/<task-id>/status.json

status values: running → complete or error

5. Read and route result

cat /Users/omarabdelmaksoud/.openclaw/workspace/comms/crewai/<task-id>/result.md
cat /Users/omarabdelmaksoud/.openclaw/workspace/comms/crewai/<task-id>/transcript.md

• Write result to relevant agent outbox (e.g. comms/outboxes/sage.md)
• Send to Vigil for quality scoring
• After Vigil approves, deliver to Omar via normal pipeline

6. Log in TASKS.json

Update the task entry:

• crewai_session: <task-id>
• transcript: comms/crewai/<task-id>/transcript.md
• status: complete

Output Files (always written, even on error)

Task Spec Format

Each --tasks argument: "Title|Description|Expected output"

• Title: Short label (for transcript)
• Description: What the agent must do (be specific)
• Expected output: What a good answer looks like (be concrete — this is what CrewAI optimizes for)

Good expected output examples:

• "A numbered list of exactly 3 risks with one-sentence explanations"
• "A Python class with docstring, __init__, and at least 2 methods"
• "APPROVED or a list of specific issues found"

Troubleshooting

Agent config not found: Run python3 ~/.openclaw/skills/crewai-collab/build_agents.py --force

Agent turn times out: Increase --turn-timeout (default 90s). Test provider: openclaw agent --agent <id> --message "ping" --json --timeout 30

Output directory already exists: Generate a new task UUID — never reuse task IDs.

hierarchical process errors: Some crewai versions require manager_agent instead of manager_llm. Fall back to sequential if hierarchical fails.

Slow run: Sequential runs are synchronous — each task waits for the prior one. For a 3-task run with 90s per turn, expect up to 5 minutes total.