multi-agent-coordination

name	multi-agent-coordination
description	Consensus protocols, conflict resolution between sub-agents, evidence merging, parallel evaluation, and delegation output grading. Use when coordinating multiple sub-agents, resolving contradictory findings, or making high-stakes decisions that warrant independent verification.

Skill: Multi-Agent Coordination

Load this when orchestrating multiple sub-agents, especially when their outputs must be merged, compared, or trusted for an important decision.

---

Core Principle

Sub-agents are specialists with fresh context. Their value is independent perspective — but independence means outputs can conflict, overlap, or vary in quality. Coordinate deliberately.

---

1. Output Grading (Confidence Scoring)

Every delegated result gets a grade before you act on it.

Grade	Criteria	Action
HIGH	Claims backed by file refs, command output, or cited sources	Trust and proceed
MEDIUM	Partial evidence, some gaps, reasonable but unverified	Verify the gap or retry narrowly
LOW	Opinion, no evidence, missing required sections	Reject — retry or escalate

Required sections in every delegation return:

• Summary (what was done/found)
• Files (paths touched or examined)
• Verification (commands run + results, or sources)
• Risks (what could be wrong)

Missing any section → automatic MEDIUM or LOW. Do not upgrade on vibes.

---

2. When to Use Parallel Consensus

Don't get consensus for everything — it's expensive. Use it only when the cost of being wrong is high.

Use consensus for:

• Irreversible actions (schema migration, production deploy, bulk delete).
• Architecture decisions with long-term lock-in.
• Conflicting initial evidence.
• Security-critical changes.

Skip consensus for:

• Reversible, low-risk edits.
• Well-understood, routine work.
• When one specialist clearly owns the domain.

---

3. Parallel Consensus Protocol

1. Frame ONE precise question.
2. Dispatch 2 agents independently (don't let them see each other's work).
   - e.g., oracle (reasoning) + jce-researcher (evidence)
3. Collect both conclusions.
4. Compare:
   - Aligned → proceed with confidence.
   - Divergent → analyze why (see conflict resolution).
   - Both uncertain → escalate to user with both views.

Independence matters: if agents share context, you get groupthink, not consensus. Give each the same question and raw inputs, not each other's answers.

---

4. Conflict Resolution

When two agents disagree:

Step 1: Identify the disagreement type
  - Factual (one is wrong about a fact) → verify the fact directly
  - Interpretive (same facts, different conclusion) → examine reasoning
  - Scope (answering different questions) → re-frame, re-dispatch

Step 2: Weigh by evidence strength, not confidence tone
  - Agent with file refs / command output > agent with assertions
  - Recent/primary sources > general knowledge

Step 3: Resolve
  - Clear evidence winner → take it, note why
  - No clear winner → run a tie-breaker check yourself
  - Genuine tradeoff → escalate to user with both sides

Never average two answers. Pick based on evidence, or escalate.

---

5. Evidence Merging

When multiple agents return complementary (not conflicting) findings:

• Deduplicate: same finding from 2 agents = 1 finding (higher confidence).
• Sequence: order findings by dependency, not by which agent returned first.
• Attribute: keep track of which evidence came from where, for traceability.
• Gap-check: what did NEITHER agent cover? That's your blind spot — investigate.

---

6. Anti-Duplication Discipline

Once delegated, do NOT redo the work yourself.

• Trust the delegated result per its grade.
• If insufficient: send a targeted follow-up for the specific gap — don't re-run the whole task.
• Never run the same search/read an agent already did.
• Two agents should never be given identical overlapping scopes (unless deliberate consensus).

---

7. Delegation Prompt Quality (6-Section Contract)

Every delegation prompt includes:

1. TASK: one-sentence atomic goal
2. EXPECTED OUTCOME: concrete deliverables + success criteria
3. REQUIRED TOOLS: explicit whitelist (Read, Grep, Bash...)
4. MUST DO: exhaustive requirements
5. MUST NOT DO: forbidden actions (modify unrelated files, skip verification)
6. CONTEXT: file paths, patterns, constraints, prior findings

Vague prompts → LOW-grade output. Invest in the prompt to avoid retry cycles.

---

8. Agent Selection Map

Need	Agent
Find files, map references, trace call paths	explorer / explore
Docs, library versions, web/GitHub evidence	jce-researcher
Hard architecture, stubborn bug, deep tradeoffs	oracle
UI, components, responsive, accessibility, visual review	frontend
Native Android build/code/release	android

Match the specialist to the work. Don't send a research question to explorer or a UI task to oracle.

---

Integration with AGENTS.md

Expands Delegation Intelligence v1.0. Load when coordinating 2+ agents or resolving conflicting outputs.

Anti-Patterns

Don't	Do
Trust delegated output blindly	Grade it (HIGH/MEDIUM/LOW) first
Get consensus for trivial decisions	Reserve consensus for high-stakes/irreversible
Let consensus agents share context	Keep them independent
Average two conflicting answers	Resolve by evidence or escalate
Redo work an agent already did	Targeted follow-up for gaps only
Send vague 1-line delegation prompts	Use the 6-section contract