Skip to content
Thoth Thoth

Multi-agent council (party)

When a single agent’s perspective isn’t enough, Thoth lets you convene a council — a structured debate of personas, each with their own mission and voice. The pattern is borrowed from the BMAD agent framework and from real human strategy councils.

It’s also the most photogenic feature Thoth ships.

What is a party

A party is a multi-agent deliberation where:

  1. N personas (typically 3–5) are spawned in parallel against the same prompt
  2. Each emits an opinion grounded in their role + voice
  3. A judge persona synthesizes the perspectives into a final response
  4. The full transcript is preserved as a Slack canvas (or thread) for review

Each persona is a separate claude -p subprocess with its own persona file. They don’t share context; they only see the original prompt and each other’s published opinions.

When to use it

  • High-stakes decisions where you want multiple frames before committing
  • Architectural choices where you want the conservative AND the bold AND the simplicity-loving voice
  • Pre-mortems where you specifically want a skeptic to find what could go wrong
  • Strategy questions where the answer needs synthesis, not retrieval

When NOT to use it

  • Routine questions — costs ~$0.50–$2 per party; not worth it for things a single Apex turn handles
  • Time-sensitive responses — parties take 30–90s due to parallel spawn + synthesis
  • Questions with a single right answer — debate is theatre when facts are clear; just ask

Default personas (BMAD-derived)

The shipped operator-agent example includes these party personas:

PersonaRoleVoice
MaryProduct strategistBig-picture, market-aware, optimistic but not naive
JohnPragmatic engineerConstraint-aware, rolls eyes at hype, cares about ops
WinstonSecurity skepticAdversarial, finds attack surfaces, slows you down (in a good way)
JamesCreative explorerGenerates options, plays with constraints, not afraid of weird ideas
Maat (judge)SynthesizerWeighs all perspectives against context, emits decision

Each persona file lives in persona/apex/party/ with the same shape as the main persona stack (SOUL, RULES, voice rules) but specialized to the role.

Invoking a party

In Slack:

/party Should we adopt agentskills.io as our skill format spec?

The bot acknowledges the party convening, then over the next ~60s:

  1. Posts each persona’s opinion as a separate Slack Block Kit Card
  2. Posts the judge’s synthesis as a final Card
  3. Optionally exports the full transcript as a Slack canvas

You can pass options:

/party Should we adopt agentskills.io as our skill format spec?
  --personas mary,john,winston      # restrict roster
  --rounds 2                        # run 2 deliberation rounds (default 1)
  --budget 2.50                     # $/cap for this party
  --canvas                          # auto-publish as Slack canvas

Cost economics

ConfigurationApprox costApprox duration
3 personas, 1 round$0.50–$0.8030–45s
4 personas, 1 round$0.70–$1.2040–60s
4 personas, 2 rounds$1.50–$2.5090–120s
5 personas + judge, 2 rounds, canvas export$2.00–$3.502–3 min

Per-party budget cap defaults to $1.50 (PARTY_DEFAULT_BUDGET_USD). Daily total cap defaults to $10 (PARTY_DAILY_CAP_USD). Beyond cap, parties are refused with a clear message.

Multi-LLM (v0.6+)

In v0.5, all personas run on Claude (homogeneous). The most-anticipated v0.6 feature is heterogeneous models per persona:

  • Mary on GPT-5
  • John on Claude Opus
  • Winston on Gemini Pro
  • James on Claude Haiku (cheap, creative)
  • Judge always on the founder’s chosen model

Research shows multi-model debates produce substantially different outcomes than single-model debates — different training data surfaces different perspectives. Plus the screenshot of three named LLMs debating in a Slack thread is unforgettable marketing.

Tracked in SPEC-multi-llm-party. Target: Q1 2027.

Recursive parties (advanced)

A persona within a party can convene a sub-party for a question that requires another deliberation. Example:

You: /party Should we ship feature X?


Mary (strategist): "Ship it. The market's moving."
John (engineer):    "Block. Our test coverage is at 40%. /party How long
                     would proper testing take?"
   └─ sub-party fires
   └─ Mary (sub): "2 weeks."
   └─ Winston (sub): "3 weeks if you want SOC2 carryover."
   └─ Maat (sub-judge): "Rule of thumb: 2.5 weeks."
John: "OK if we delay 2.5 weeks for proper coverage, ship. Otherwise no."
Winston: "Plus a security review pass."
Maat (top-judge): "Ship in 3 weeks with covered tests + security pass."

Recursive parties are bounded by depth (default max 2 levels) and budget (sub-parties debit from the same envelope).

How parties influence the rest of Thoth

A party transcript is first-class data:

  1. The full debate is stored in bridge.db parties + party_messages tables
  2. Reactions on individual persona Cards influence reflection signal (e.g., ✅ on Mary’s opinion → Mary’s reasoning gets weighted higher in skill compilation)
  3. Synthesized decisions written to MEMORY.md become durable references
  4. The dashboard’s Council tab visualizes party transcripts in a ritual-scroll format

Disabling

Terminal window
PARTY_DISABLED=true     # disables /party command entirely

You can also disable just specific personas by removing their files from persona/apex/party/.

What’s next