decide

AgentCouncil Decide

You are running the AgentCouncil decide protocol. You are the orchestrator — you define the decision space and provide context. The outside agent is the independent evaluator.

Decision: $ARGUMENTS

Backend Selection

Parse the arguments for backend=<value> (e.g., backend=codex, backend=claude, backend=ollama). If omitted, the default backend is used. 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 argument (or omit profile if no backend= argument was given).

Save workspace_access from the response.

Step 1: Identify the decision and define options

From the conversation context and codebase, determine:

• The decision question
• The options (minimum 2, each with id, label, description)
• Evaluation criteria (what matters?)
• Constraints (what limits the choice?)

Display the decision space so the user can see exactly what's being evaluated.

Step 2: Send decision request to outside agent

Construct an evaluation request that gives the evaluator:

• The decision question
• Each option with clear descriptions
• Criteria and constraints
• Explicit instruction: evaluate ONLY these options, do NOT invent new ones

For file context:

• If workspace_access is native: include file paths for the evaluator to read directly. Do NOT dump file contents. Add this hint at the end of the evaluation request: "If you have access to code navigation tools (e.g. serena, codegraph), use them to understand the codebase before evaluating options."
• If workspace_access is assisted or none: read the relevant files yourself and include key content in the prompt.

Call mcp__agentcouncil__outside_start with prompt set to the evaluation request and profile set to the backend argument (or omit profile for default). Save session_id from the response. Save response as the evaluator's first reply.

Step 3: Read assessments and respond with context

The evaluator returns: for each option, pros, cons, blocking risks, disposition, confidence.

Read the assessments and respond with your codebase knowledge:

• Correct factual errors: "Option B's 'extra infra' con is overstated — we already run Redis"
• Add context: "Option A integrates with our existing auth middleware, see auth.py"
• Challenge assumptions: "The latency claim for Option C assumes us-east — we're in eu-west"

Send your response with context corrections via mcp__agentcouncil__outside_reply with the saved session_id. Include the synthesis request prompt:

"Based on my context corrections, produce a final decision synthesis. Return JSON with these exact keys: outcome (decided/deferred/experiment), winner_option_id (string or null), decision_summary (string), option_assessments (array of objects with option_id, pros, cons, blocking_risks, disposition, confidence, source_refs), defer_reason (string or null), experiment_plan (string or null), revisit_triggers (array of strings), next_action (string). Rules: evaluate ONLY the provided options. For EVERY option include assumptions, tradeoffs, and confidence. Set disposition to selected/viable/rejected/insufficient_information. Outcome: decided if one option clearly superior, deferred if insufficient info, experiment if needs real-world validation."

Step 3.5: Close session

Call mcp__agentcouncil__outside_close with the saved session_id.

Step 4: Present the structured result

Parse the JSON response from Step 3 and present:

## Decision Result

**Outcome:** {outcome}
**Winner:** {winner_option_id or "None"}

### Decision Summary
{decision_summary}

### Option Assessments
For each option:
- **{option_id}: {label}** (disposition: {disposition}, confidence: {confidence})
  Pros: {pros}
  Cons: {cons}
  Blocking Risks: {blocking_risks or "None"}

### Revisit Triggers
- {each trigger}

### Next Action
{next_action}

If JSON fails to parse, report outcome as "deferred" and show raw response.

Rules

• You define the decision space — the evaluator works within it
• Do NOT write your own assessment before sending to the outside agent
• The evaluator can ONLY evaluate provided options — it cannot invent new ones
• Always call outside_close after the final synthesis response
• The session layer handles history management — just send the new message each turn
• Adapt file inclusion to workspace_access: file paths for native, inline contents for assisted/none