| name | council-session |
| description | Multi-model consensus via subagent DAG. Use when you need diverse model perspectives on complex architectural decisions, high-stakes code changes, or ambiguous trade-offs. |
Council Session
Run a multi-model council to get diverse AI perspectives on a complex problem.
When to Use
- Critical architectural decisions needing diverse model perspectives
- High-stakes code changes where consensus reduces risk
- Ambiguous problems where multi-model disagreement is informative
- Security-sensitive design reviews
When NOT to Use
- Straightforward tasks you're confident about
- Speed matters more than confidence
- Routine implementation work
Pre-built Agents
The following agents are pre-configured with different models in ~/.kiro/agents/:
| Agent | Model | Role |
|---|
councillor-a | claude-opus-4.6 | Strong reasoning, deep analysis |
councillor-b | glm-5 | Different training data, different biases |
councillor-c | claude-opus-4.5 | Balanced reasoning, prior generation perspective |
council-master | claude-opus-4.6 | Synthesizes all councillor outputs |
All councillors are read-only (can read files, grep, glob but cannot edit). The council-master is pure synthesis.
To change models, re-run ~/.kiro/generate-configs.sh after editing the model values in the script.
How to Run a Council
You (code_supervisor) construct a 4-stage subagent call:
- 3 parallel councillor stages (no dependencies — they run simultaneously)
- 1 council-master stage that
depends_on all 3 councillors
Example Subagent Call
{
"task": "<describe the problem/decision here>",
"stages": [
{
"name": "councillor-a",
"role": "councillor-a",
"prompt_template": "{task}"
},
{
"name": "councillor-b",
"role": "councillor-b",
"prompt_template": "{task}"
},
{
"name": "councillor-c",
"role": "councillor-c",
"prompt_template": "{task}"
},
{
"name": "council-master",
"role": "council-master",
"depends_on": ["councillor-a", "councillor-b", "councillor-c"],
"prompt_template": "Synthesize the optimal response for:\n{task}"
}
]
}
Each councillor's system prompt is already defined in its .md file. The prompt_template only needs the problem description — use {task} to pass through the overall task.
Result Handling
Present the council master's synthesized response verbatim to the user. Do not re-summarize — the council master has already produced the final answer.
