multi-agent-orchestrator

name	multi-agent-orchestrator
description	Defines DAG-style multi-skill runs with explicit input/output paths, approval gates between steps, failure rollback notes, and a run log markdown—writes to workspace/drafts/orchestration/; complements social-media-manager for cross-domain chains.
metadata	{"clawdbot":{"emoji":"🧩"},"openclaw":{"emoji":"🧩"}}

multi-agent-orchestrator

Meta-orchestration for SkillsToAdd.md chains: explicit DAG, files in / files out, gates—same clarity as qf-course-researcher step list (web → API → Notion rows).

Execution note: Skills are not auto-invoked by the filesystem. “Wiring” = shared paths (README-handoff.md, calendar.md, pipeline-state.md, APPROVAL.md) + you (or OpenClaw cron) running the next skill. The canonical social DAG lives in chain-templates/social-feed-pipeline.md; social-media-manager lists the same order in pipeline-state.md.

Automatic handoffs — what OpenClaw can (and cannot) do

Mechanism	What it does	Feels like “auto handoff”?
Built-in skill chaining	None. Finishing marketer-agent does not enqueue social-media-manager.	No
Cron (openclaw cron add)	Gateway wakes on a schedule and runs an isolated or main session with your --message. That message can say: read pipeline-state.md, run the next incomplete step, update the file.	Yes — if you encode “next step” in the prompt
Webhooks (hooks in config)	POST /hooks/agent or /hooks/wake starts a turn (e.g. CI or n8n after a file lands in Drive).	Yes — event-driven
Custom session (--session session:social-pipeline)	Same session id across recurring cron runs so the agent remembers prior RUNLOG / pipeline state (within retention).	Yes — for multi-day chains
One giant cron message	“Do steps 3→4→5 in one turn.”	Possible but fragile (timeouts, gates, token limits); prefer one step per job or explicit checklist in one message

Practical pattern for your social pipeline

1. Cron job A (e.g. weekly): --message = Open workspace/drafts/social/<campaign>/pipeline-state.md. If calendar is incomplete, run social-media-manager + social-content-planning; else if post_bodies incomplete, run writers for missing rows; append RUNLOG.
2. Cron job B + C (AM/PM): Only send approved rows to hype-engine (POST /posts with date/time, empty UUID filter per slot time)—see social-media-manager cron examples; no separate publish-now step.
3. Keep APPROVAL.md as the human gate so cron does not post without approved rows.

Docs: Scheduled tasks (cron) · Webhooks (same page) · openclaw cron list / openclaw cron add --help

Prerequisites

• Human: goal, deadline, risk tolerance (can any step auto-execute?).
• List of available skills in workspace skills/ (read names from folders).
• Output:

  workspace/drafts/orchestration/<YYYY-MM-DD>-<run-slug>/
  

Credentials & API (qf-style)

• Orchestration docs only: The DAG does not require secrets by itself.
• Per-step needs: When a step lists “Tool/API”, resolve credentials via workspace/INTEGRATIONS.md and the target skill’s “Credentials & API” section (qf-course-researcher, notion, adverts-creator, etc.).

High-level Workflow

1. Run charter (00-charter.md)

   • Objective, success metric, out of scope, owner human.

2. DAG (dag.mmd or dag.md)

   • Mermaid or numbered list: Step N: skill → reads → writes.

3. Step specs (steps/)

   • One file per step 01-<skill>.md:
     • Inputs: exact paths or “TO BE PRODUCED BY STEP N-1”
     • Outputs: exact paths
     • Tool/API needs
     • Approval gate: none | human | human+legal

4. Recommended chains (templates)

   • chain-research-pm.md — auto-research-agent → product-manager-agent (or product-manager)
   • chain-research-marketer.md — auto-research-agent → marketer-agent
   • chain-marketer-social.md — marketer-agent → social-media-manager → optional adverts-creator
   • chain-pm-marketer.md — product-manager-agent → marketer-agent (launch messaging after PRD)
   • chain-agency-marketing-execution.md — agency-marketing (persona) → marketer-agent → social-media-manager
   • chain-agency-pm-delivery.md — agency-project-management (shepherd or senior PM) → product-manager-agent (PRD) as needed
   • chain-trend-tiktok.md — social-trend-monitor → tiktok-video-ads-creator
   • chain-image-ads.md — auto-image-generation → adverts-creator
   • chain-li-social.md — linkedin-article-writer → social-media-manager
   • chain-social-feed-full.md — marketer-agent → social-media-manager (calendar) → post bodies → auto-image-generation (per slot, brand-images/) → optional social-caption-writer → bundles → hype-engine (Media + scheduled POST /posts, UUID idempotency). Concrete file: chain-templates/social-feed-pipeline.md.
   • chain-video-tiktok.md — auto-video-generation → tiktok-video-ads-creator
   • QuizFactor: reference existing qf-* instead of duplicating.

5. Run log (RUNLOG.md)

   • Append-only: timestamp, step, status, output path, errors.

6. Rollback (rollback.md)

   • If step K fails: which files are stale; what to re-run.

7. Completion message

   • User: start at steps/01-... and follow order; stop at gates.

8. Scheduling

   • Cron-style note in charter if daily/weekly — use openclaw cron add (see Automatic handoffs above), not a fictional built-in chain.

Outputs (required)

• 00-charter.md, dag.md, steps/ directory with ≥1 step file, RUNLOG.md, README-handoff.md

Agent Checklist

• Every step has readable inputs (not vague “data”).
• Every output path is under workspace/drafts/... or explicit tool destination.
• Gates labeled; destructive steps flagged.
• qf-* chains not re-specified in full—point to existing skills.
• RUNLOG template initialized with run id.
• User instructed to execute steps sequentially unless parallel safe.