CtrlK
BlogDocsLog inGet started
Tessl Logo

agent-orchestration

Multi-agent orchestration: workflow execution, triad composition, routing, parallelism, subagent monitoring, error recovery, and learning loops. Topics: continuous-learning, error-recovery-automation, intelligent-routing, multi-model-routing, parallel-workflow, socratic-debate, subagent-monitor, task-automation, triad-composition, workflow-orchestration.

SKILL.md
Quality
Evals
Security

agent-orchestration

Router skill: dispatches multi-agent orchestration work to exactly one playbook. [DOC]

When to use

Any request to compose/route/run/recover/monitor a multi-agent workflow under PRISTINO. [DOC] NOT for single-agent tasks, content generation, or domain skills — route those to their own skill. [INFERENCE]

Contract

  • Input: topic (one of the 10 enum values; infer from request, ask only if two routes tie) + depth. [CONFIG]
  • Output: the resolved playbook applied; every non-obvious claim tagged from the Alfa core set ([CODE]/[CONFIG]/[DOC]/[INFERENCE]/[ASSUMPTION]) — never mix tag families. [DOC]

Procedure

  1. Resolve topic from the request. Read EXACTLY ONE playbook from routes:. Never load the cluster. [CONFIG]
  2. depth=deep → apply exhaustively, verify at each step; quick → essentials only. Default quick. [CONFIG]
  3. Run the spine: Discover → Analyze → Execute → Validate. [DOC]
  4. Gate before "done": constitution v6.0.0 enforcement, evidence tags present, script-first rule honored. [CONFIG]

Topic disambiguation (pick the narrowest match)

  • Role/Lead-Support-Guardian selection → triad-composition; which model per task → multi-model-routing. [DOC]
  • Which skill/agent handles a request → intelligent-routing; end-to-end run → workflow-orchestration. [DOC]
  • Concurrent fan-out → parallel-workflow; live subagent status → subagent-monitor. [DOC]
  • Multi-perspective reasoning → socratic-debate; harvesting its insights → continuous-learning. [DOC]
  • Recover from a failed step → error-recovery-automation; schedule/trigger work → task-automation. [DOC]

Acceptance gate

Done only if: one playbook loaded (not many); spine completed; gates passed; tags single-family + EN/ES consistent. [DOC] Walk assets/routing-checklist.md before emitting; Guardian scores with assets/quality-rubric.json. [CONFIG]

Anti-patterns

  • Loading multiple playbooks "to be safe" — defeats the router; pick one or ask. [INFERENCE]
  • Guessing topic when two routes genuinely tie — ask instead of mis-routing. [ASSUMPTION]
  • Marking complete without the constitution/evidence/script-first gate. [DOC]
  • Inventing a topic outside the enum — the catalog parser rejects it. [CONFIG]

Self-correction

If mid-run the request no longer fits the chosen topic, stop, re-resolve against the disambiguation table, and switch playbooks — do not stretch the wrong one. [INFERENCE]

Repository
JaviMontano/jm-adk-beta
Last updated
First committed

Is this your skill?

If you maintain this skill, you can claim it as your own. Once claimed, you can manage eval scenarios, bundle related skills, attach documentation or rules, and ensure cross-agent compatibility.