# Agent Council > Run a bounded discussion between the current host agent, Grok Build CLI, and Claude Code CLI. Use when the user requests a council, multi-agent debate, or 3–5 rounds of discussion before a recommendation. Bundled Python scripts coordinate full shared history, exact native sessions, host contributions, final assessments, and crash recovery. The current agent participates; never spawn another host or invent its turns. - Skill: `gordon1210/agent-council` (Agent Skill, multi-file: 13 files) - Install (CLI): `npx skillmds@latest add gordon1210/agent-council` - Raw SKILL.md: https://api.skillmd.com/api/skills/gordon1210/agent-council/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: AI & ML - License: MIT - Author: gordon1210 (https://skillmd.com/u/gordon1210) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/gordon1210/agent-council --- # Agent Council You are the **host participant and moderator**. Ask Grok and Claude to discuss an authorized topic with you, contribute your own reasoning each round, then present a verified synthesis. Use the bundled scripts; do not improvise another session manager, parse terminal UI text, or launch a second copy of yourself. ## Discussion contract Default: **four rounds**, each containing one host, one Grok, and one Claude contribution in that order. Three to five rounds is the normal range. A round is not a provider's `--max-turns`, which bounds its internal agent loop. Every peer call receives the topic, supplied context, approved updates, and **every previous discussion contribution verbatim** in one canonical snapshot. The default also resumes that peer's exact native session. Both peers receive a final assessment call against the same completed discussion snapshot. You read the full discussion plus both assessments before writing the result. Four rounds therefore use **10 peer CLI calls**: eight discussion calls and two final assessments. The current host's contributions and synthesis do not start extra CLIs. Final assessments are separate from discussion turns; they are not recursively broadcast to each other. This avoids an endless final-sync loop. Never claim a peer read a contribution written after its input snapshot. ## Scope and prerequisites Sending material to both providers must be authorized. Include selected source excerpts, constraints, and verified facts; never forward your whole host chat, private hidden reasoning, credentials, or unrelated files. Discussion is **tool-free**: no implementation, shell, web, MCP, or nested agents. Gather any needed evidence yourself within your existing authority and supply it as context. Resolve this skill's actual installed directory and set `COUNCIL` to its `scripts/council.py` path. Run `doctor` in the intended project: ```bash python3 "$COUNCIL" doctor --cwd /absolute/project ``` Review its CLI version/help and Grok `inspect` output locally. Review applicable Claude managed policy as well. Neither empty built-in tools nor a permission mode disables all possible startup hooks/plugins/MCP processes. Do not proceed unless remaining executable behavior is disabled or reviewed as compatible with this consultation. `--config-reviewed` below attests to that review; it is not proof generated by the script. Do not send private preflight output to peers. The narrow adapters follow the adjacent **grok-cli** and **claude-cli** skills' authority boundaries. Read those skills for broader CLI diagnostics if installed; the council's scripts are nevertheless self-contained and do not import sibling skills. No CLI installation, update, login, config mutation, or credential proxy is performed. See [provider contracts](references/provider-contracts.md). ## Run the council Prepare a UTF-8 topic packet and optional evidence files. Initialize a **new** run directory in an approved location, preferably outside the repository: ```bash python3 "$COUNCIL" init \ --run-dir /approved/work/council-01 \ --cwd /absolute/project \ --topic-file /approved/work/topic.md \ --context-file /approved/work/evidence.md \ --rounds 4 --config-reviewed ``` Omit `--context-file` if unnecessary; repeat it for multiple selected files. No model is called by `doctor` or `init`. Repeat this host-driven loop until `awaiting-synthesis`: 1. Run `brief --run-dir ...`. **Read the complete returned snapshot**, not just status or the last response. It includes a `revision` and the next speaker. 2. At `awaiting-host`, write your own substantive contribution to a UTF-8 file. Use `host --run-dir ... --message-file ... --expect-revision N`, replacing `N` with the exact revision just read. Never manufacture a host contribution with another CLI or append a placeholder just to let automation continue. 3. Run `advance --run-dir ...`. It calls only the due peers, persists each accepted contribution, and stops at the next host turn or synthesis boundary. On the last round it also performs the two final assessments. 4. Read `brief` again. Answer concrete objections and cite message IDs. Do not merely summarize peer answers: the host is the third participant. At `awaiting-synthesis`, read the complete `brief`, independently check material claims as needed, and write the final synthesis to a file. Call: ```bash python3 "$COUNCIL" finalize \ --run-dir /approved/work/council-01 \ --message-file /approved/work/synthesis.md \ --expect-revision N ``` The script writes `report.md`; `transcript.md`, `transcript.jsonl`, native session IDs, requests, outputs, and recovery state remain available. Present the actual result to the user, not just an artifact path. Do not present a complete council unless the run actually reached the synthesis boundary. ## Quality and stop rules Early turns establish alternatives and assumptions. Middle turns challenge specific claims and update positions. The final discussion round resolves the strongest objections and records remaining blockers. Final synthesis separates recommendation, evidence-backed agreement, unresolved dissent, and next checks. A majority opinion is not evidence. No manufactured consensus or model-brand stereotypes. Preserve meaningful disagreement even when all other agents agree. Use `note` with a fresh revision for an approved correction or additional evidence during discussion; never edit old contributions. Stop on provider failure, denied permissions, unknown result shape, missing/mismatched session ID, timeout, or context-budget overflow. Do not silently truncate, summarize away history, skip a peer, change model, broaden tools, or drop a sandbox. `pending-recovery` requires inspection, then either accepting an already valid captured result or explicitly discarding the attempt. Discarding abandons that peer's possibly contaminated native session; the next call reconstructs context from the full canonical transcript. Follow [recovery and privacy](references/recovery-and-privacy.md), not ad-hoc retries. For early user cancellation or unavailable peers, `finalize --partial --reason` is available after resolving any pending attempt. Label the report **partial** and identify missing input. Never present it as the completed requested debate. See [workflow](references/workflow.md) for commands and rationale, [provider contracts](references/provider-contracts.md) for boundaries, and [validation](references/validation.md) for tests and evaluation scenarios.