brainstorm

AgentCouncil Brainstorm

You are running the AgentCouncil deliberation protocol. You are BOTH the orchestrator and the lead agent. The outside agent runs independently.

Topic: $ARGUMENTS

Backend Selection

Parse the arguments for backend=<value> (e.g., backend=codex, backend=claude, backend=ollama). If omitted, the default backend is used. The backend argument selects which provider to use. Provider differences (session persistence, workspace access) are handled automatically by the session layer.

Protocol — follow these steps exactly

Step 0: Get backend capabilities

Call mcp__agentcouncil__get_outside_backend_info with profile set to the parsed backend value. If no backend= was specified, omit the profile argument.

Save workspace_access from the response. This determines how to construct prompts:

• If workspace_access is "native": reference file paths in prompts — the outside agent can read them directly.
• If workspace_access is "assisted" or "none": read relevant files yourself and inline their contents in prompts — the outside agent cannot access the workspace.

Step 1: State the problem

Clearly state what needs to be decided and why it matters. This frames the deliberation for both agents and the user.

Step 2: Create the brief and send to outside agent FIRST

Write a neutral brief containing ONLY:

• Problem statement
• Background context
• Constraints
• Goals
• Open questions

CRITICAL: Do NOT include any proposal, opinion, or direction. The brief must be clean.

For prompt construction:

• If workspace_access is "native": include file paths in the brief for the outside agent to read. Add this hint at the end of the brief: "If you have access to code navigation tools (e.g. serena, codegraph), use them to understand the codebase structure before forming your proposal."
• If workspace_access is "assisted" or "none": read the relevant files yourself and include their contents in the brief.

Call mcp__agentcouncil__outside_start with prompt set to the brief text, profile set to the backend argument (or omit profile for default), and await_response set to false. This fires the brief to the outside agent and returns immediately with session_id and status: "pending". The outside agent is now working on its proposal in the background.

Step 3: Write YOUR proposal (parallel with outside agent)

The outside agent is processing the brief in the background. Write your own independent proposal NOW — do not wait. You have the full conversation context — use it. Be specific and opinionated.

Display it clearly labeled as Lead agent's proposal.

Independence is preserved because the brief was neutral (no opinion) and you have not read the outside agent's response yet. Both agents think independently.

Step 4: Read outside agent's proposal and share yours

Call mcp__agentcouncil__outside_read with the saved session_id to fetch the outside agent's response (this blocks until ready, but the agent has been working since Step 2).

Display it clearly labeled as Outside agent's proposal.

Then send your full proposal so the outside agent can compare both views.

Call mcp__agentcouncil__outside_reply with the saved session_id and this prompt:

"Here is the lead agent's independent proposal (written before seeing yours): {your full proposal from Step 3}. Compare it with your own proposal. Where do you agree? Where do you disagree? Push back where you think the lead is wrong."

Step 5: Compare proposals

Now both agents have seen each other's work. Display the outside agent's comparison response. Summarize where the two proposals agree and disagree.

Step 6: Exchange rounds (with early exit)

Default to 1 round unless the user asks for more. For each exchange after Step 5:

Read the outside agent's response. If there are no material disagreements — both proposals align on direction, details, and tradeoffs — skip remaining exchanges and go directly to Step 7 (synthesis). State: "No material disagreements — proceeding to synthesis."

If disagreements exist: Write your counter-response addressing their pushback.

Call mcp__agentcouncil__outside_reply with the saved session_id and the new prompt. The session layer handles history management — just send the new message each turn.

Step 7: Get synthesis

Call mcp__agentcouncil__outside_reply with the saved session_id and this prompt:

"Based on our full discussion, produce a final synthesis. Return JSON with these exact keys: recommended_direction (string), agreement_points (list of strings), disagreement_points (list of strings), rejected_alternatives (list of strings), open_risks (list of strings), next_action (string), status (one of: consensus, consensus_with_reservations, unresolved_disagreement). Do NOT use partial_failure as status."

Step 8: Present the result

Parse the JSON response and present:

## Brainstorm Result

**Status:** {status}

### Recommended Direction
{recommended_direction}

### Agreement Points
- {each point}

### Disagreement Points
- {each point}

### Rejected Alternatives
- {each point}

### Open Risks
- {each risk}

### Next Action
{next_action}

If JSON fails to parse, report status as "unresolved_disagreement" and show raw response.

Step 8.5: Close session

Call mcp__agentcouncil__outside_close with the saved session_id.

Rules

• Default to 1 round unless the user asks for more
• ALWAYS send the brief FIRST so the outside agent starts immediately
• Write your proposal AFTER sending the brief but BEFORE reading the response
• NEVER put your opinion in the brief
• Always call outside_close after the final synthesis response
• The session layer handles history management — just send the new message each turn
• Adapt prompt construction to workspace_access: file paths for native, inline contents for assisted/none
• Do NOT add analysis beyond the structured result