Reunion de Agentes

Reunion de Agentes

Purpose: Structured alignment between 2-5 agents → concrete decisions, not transcripts.

Output: Decision doc at docs/meetings/YYYY-MM-DD-{topic}.md

Core principle: The output is for HUMANS. If a human can't read it in 2 minutes, it's too long.

Protocol

Step 0: Setup

Infer or ask:

1. Tema — What needs alignment?
2. Participantes — Which agents?
3. Decisiones esperadas — What concrete decisions must come out?

Step 1: Create Meeting File

Create docs/meetings/YYYY-MM-DD-{topic-slug}.md:

# {Tema}

> **Fecha:** YYYY-MM-DD | **Participantes:** Agent1, Agent2, ... | **Issue:** SKY-XXX

## Decisiones necesarias

1. {decision 1}
2. {decision 2}
...

---

## Propuestas

### {Agent1}

| Decisión | Propuesta | Justificación |
|---|---|---|
| {decisión 1} | {respuesta concreta} | {1 línea por qué} |
| {decisión 2} | {respuesta concreta} | {1 línea por qué} |

**Necesito de otros:** {1-2 bullets max}
**Conflicto que veo:** {1 bullet max, o "ninguno"}

### {Agent2}
...

---

## Conflictos

| # | Tema | {Agent1} dice | {Agent2} dice | Resolución |
|---|---|---|---|---|
| 1 | ... | ... | ... | _pendiente_ |

---

## Decisiones finales

| # | Decisión | Resolución | Owner |
|---|---|---|---|
| 1 | ... | ... | ... |

## Timeline

| Día | Entregable | Owner | Depende de |
|---|---|---|---|
| D+0 | ... | ... | — |

## Pendientes (usuario)

- ...

Step 2: Propuestas (PARALLEL)

Launch ALL agents in parallel. Each agent:

• Reads its definition at hive/agents/{agent}.md
• Reads the meeting file
• Fills its table under "Propuestas"

CRITICAL format rules for agents:

You are {AgentName} for Skywalking.dev.
Read your definition at hive/agents/{agent-file}.md
Read the meeting file at {path}.

Fill your "Propuestas" table. Rules:
- ONE ROW per decision needed. Answer each decision directly.
- "Propuesta" column: concrete answer (hex code, lib name, file path, number). NO prose.
- "Justificación" column: ONE sentence max. Why this choice.
- "Necesito de otros": max 2 bullets. What blocks you.
- "Conflicto que veo": max 1 bullet. What will clash with another agent. Or "ninguno".
- TOTAL max 15 lines. If you write more, you failed.
- Spanish. No filler. No "estoy de acuerdo con la visión de...". Just answers.

Step 3: Conflict Resolution (PARALLEL, only if needed)

After reading all proposals, Mentat:

1. Fills the "Conflictos" table — where agents disagree
2. If conflicts exist (>0 rows), launches ONLY the conflicting agents in parallel with:

You are {AgentName}. Read the meeting file at {path}.
You disagree with {OtherAgent} on: {conflict description}.
In the Conflictos table, write your final position in your column.
ONE sentence. Include what you'd accept as compromise.

3. If no conflicts, skip to Step 4.

Max 1 conflict round. If unresolved, Mentat decides.

Step 4: Resoluciones (MENTAT)

Mentat reads all proposals + conflict resolutions and writes:

1. Decisiones finales — table with resolution + owner per decision
2. Timeline — table with day, deliverable, owner, dependency
3. Pendientes (usuario) — questions only the user can answer

That's it. No "acuerdos alcanzados", no "conflictos resueltos" recap, no redundant sections. The tables ARE the decisions.

Step 5: Present to User

Show the Decisiones finales table + Pendientes. Link to meeting file. Done.

Format Rules (NON-NEGOTIABLE)

• Tables > prose. Every decision is a table row, not a paragraph.
• 15 lines max per agent. Agents that write more get their output truncated.
• No repetition. If Agent1 said X in Propuestas, Agent2 doesn't restate it in Conflictos.
• No "estoy de acuerdo". Agreement = same value in the table. No need to narrate it.
• No preamble. Agents don't explain their role, restate the context, or summarize what others said.
• Concrete values only. "warm earth tones" is not a proposal. #C4A86A, #8B7355, #7A9B7F is.
• Spanish. ES-LATAM, terse.
• Total meeting file < 100 lines. If it exceeds this, the format failed.

Agent Reference

Agent	subagent_type	Domain
Aurora	aurora-product-designer	Brand, visual design, UI specs
Kokoro	kokoro-backend-specialist	APIs, databases, auth, security
Pixel	pixel-frontend-specialist	Next.js, React, UI implementation
Flux	flux-automation-specialist	n8n workflows, integrations
Hermes	hermes-deployment-specialist	Vercel, CI/CD, performance
Centinela	centinela-test-automation-expert	Testing, QA, Playwright
Lumen	lumen-seo-specialist	SEO, visibility, schema markup
Pregon	pregon-marketing-specialist	Content, social, campaigns
Oraculo	oraculo-research-specialist	Web research, competitive intel

Note: For agents without a dedicated subagent_type, use general-purpose with their definition from hive/agents/{name}.md.

Anti-patterns

• No transcripts. This is a decision doc, not meeting minutes.
• No "Ronda 1 / Ronda 2" format. Proposals + conflicts + decisions. Three sections.
• No agent monologues. Tables enforce structure.
• No Mentat opinion in Propuestas. Mentat resolves conflicts, doesn't propose.
• No meetings for 1 agent. Direct delegation instead.
• No meetings without decision targets. If nothing needs deciding, don't meet.