Converge

Structured debate and refinement using Agent Teams. Takes divergent findings (from /diverge or any independent research) and spawns a team where teammates advocate for different positions, debate trade-offs, and converge on a refined synthesis.

Arguments

$ARGUMENTS — format: [N] [path/to/diverge_output.md or inline topic] where N is optional number of debaters (default: 3, min 2, max 5)

Prerequisites

• Agent Teams must be enabled (CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1)
• Input should be divergent findings — either a file path to prior /diverge output, a PRD with open tensions, or a description of competing approaches

Parse Arguments

Extract:

• team_size: optional leading integer (default 3, min 2, max 5)
• input: path to a file with divergent findings, or inline description of the debate topic

Phase 0: Validate Teams Availability

Check that Agent Teams is enabled. If not, inform the user:

Agent Teams is required for /converge. Enable it by adding to your Claude Code settings:
  "env": { "CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS": "1" }
Or use /diverge for independent exploration without debate.

Phase 1: Frame the Debate

If a file path is provided: read the file and extract the key tensions, divergence points, and competing recommendations.

If inline input: parse the competing positions from the description.

If neither is clear: ask the user what positions or findings should be debated.

Prepare a debate brief that includes:

• The core question or decision to be made
• 2-5 distinct positions to be advocated (one per teammate)
• Shared context and constraints all positions must respect
• Evaluation criteria: what makes one approach better than another?

Present the debate brief and position assignments to the user for confirmation.

Phase 2: Spawn the Team

Create an agent team with the following structure:

Team Lead (you): Moderator and synthesizer. You frame rounds, call for responses, and produce the final synthesis.

Teammates: Each assigned one position to advocate for. Their instructions:

You are a debate participant in a structured convergence exercise.

## Debate Brief
{debate_brief}

## Your Position: {position_name}
{position_description}

## Your Role
You are the advocate for this position. Your job is to:
1. Make the strongest possible case for your position
2. Identify weaknesses in other positions when they're shared
3. Acknowledge genuine strengths of competing positions (steel-man, don't straw-man)
4. Propose concrete compromises when positions partially overlap
5. Be willing to update your view if presented with compelling arguments

## Communication Rules
- When you receive a message from another teammate, respond substantively
- Challenge weak arguments but acknowledge strong ones
- If you genuinely think another position is better on a specific dimension, say so
- Focus on trade-offs, not absolutism — most positions have merit in some context

## Output Format for Each Round
- **Claim**: Your key argument this round (1-2 sentences)
- **Evidence**: Supporting reasoning or data
- **Concession**: What the other positions get right
- **Challenge**: Specific weakness in a competing position
- **Synthesis Offer**: How your position could incorporate the best of others

Phase 3: Run Debate Rounds (2-3 rounds)

Round 1: Opening Positions

Each teammate presents their strongest case. As lead, broadcast all opening positions so everyone can see them.

Round 2: Challenges and Responses

Each teammate responds to the positions they find most compelling or most flawed. Direct them to engage with specific claims from Round 1.

Round 3: Synthesis Proposals (optional — run if still divergent)

Each teammate proposes their best "combined" approach that incorporates insights from the debate. This is where compromise positions emerge.

Moderation rules:

• If a round produces consensus on a point, mark it as resolved and move on
• If a round surfaces a new tension, flag it for the next round
• Keep rounds focused — redirect tangents back to the evaluation criteria
• After Round 2, assess whether Round 3 is needed (skip if positions have already converged)

Phase 4: Synthesize

After debate concludes, produce the convergence synthesis:

Convergence Report Structure

1. Resolved Points Positions where the team reached consensus through debate. Note what argument or evidence was decisive.

2. Refined Trade-offs Tensions that weren't resolved but are now better understood. For each:

• The trade-off in precise terms
• What context or constraints would tip the balance one way vs the other
• What information would be needed to resolve it

3. Emerged Positions New approaches that didn't exist before the debate — combinations or compromises that teammates proposed during synthesis rounds.

4. Strongest Arguments The single most compelling argument from each position, preserved even if the overall position wasn't adopted.

5. Recommended Path A concrete recommendation that:

• Incorporates resolved consensus points
• Makes explicit choices on unresolved trade-offs (with rationale)
• Flags which choices should be revisited with more data

6. Debate Highlights Per-teammate: the single strongest contribution they made to the discussion.

Phase 5: Update Artifacts

If the debate was based on a /diverge PRD:

• Update the PRD with refined requirements based on debate outcomes
• Mark resolved open questions as resolved
• Add new considerations surfaced during debate
• Save updated PRD and present path to user

If no PRD exists, offer to draft one from the convergence synthesis.

Rules

• Steel-man, don't straw-man: teammates must engage with the strongest version of each position
• The lead moderates, not advocates: as team lead, stay neutral during rounds. Only take a position in the final synthesis.
• Preserve dissent: if a teammate remains unconvinced, include their dissent in the final report. Forced consensus is worse than honest disagreement.
• 2-3 rounds max: debates have diminishing returns. If positions haven't moved by Round 3, synthesize with the tensions acknowledged.
• Evaluation criteria are fixed: don't let the debate shift the goalposts. If criteria need updating, that's a finding, not a mid-debate change.
• Token awareness: Teams are expensive. Keep rounds focused and don't let teammates repeat themselves.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: sjarmak
• Source: sjarmak/agent-workflows
• License: MIT

Install and usage instructions live in the source repository linked above.