Daniel StoyanovQuorumOpenAI Agents SDK · Python

management agent · advisory panel

Quorum

Four perspectives, one recommendation

Agenda Overview
OpenAI Agents SDK · Python
Multi-agent handoffs with a synthesising moderator

An advisory panel for contested calls. A triage agent routes the question to four perspectives that argue from genuinely opposed viewpoints (growth, risk, engineering, customer), and a moderator turns them into a recommendation with the dissenting view kept beside it, not smoothed over. The decision stays with you.

Take a look inside

3 questions inside

Item 1 · How the panel works

What makes it special

  1. Four perspectives, four agents

    The growth, risk, engineering and customer perspectives each run as a separate agent with its own brief, not one model role-playing four voices. The question passes by real handoff with a stated reason, and every perspective starts from a context stripped of the others' tool traffic.

    Source: quorum/board.py · handoff(), remove_all_tools

  2. The dissenting view stays in the advice

    The moderator has to recommend something (“it depends” is not an output) and may advise against even the most confident perspective. The dissenting view is a required part of every recommendation, so you see who disagreed and why instead of a split smoothed into consensus. The decision stays with you.

    Source: quorum/models.py · Decision.dissent, required

  3. A checker that sends advice back

    The panel’s checks are enforced in code, not requested in a prompt. A recommendation with no falsifier - the evidence that would reverse it - is rejected and returned to the moderator, and so is any recommendation attempted before all four perspectives have stated.

    Source: quorum/record.py · board_rules.decision

  4. Out of scope, never discussed

    A scope guardrail checks the question before any perspective is engaged. Off-remit asks, such as personal legal advice, are declined with a referral note: zero perspectives heard, zero handoffs.

    Source: quorum/guardrails.py · keyword rules: predictable, blind to paraphrase

Item 2 · Questions put to the panel

Inside the demo

  1. Question 01

    Sunset the free tier?

    Four perspectives argue a pricing question. Watch the checker send back a recommendation with no falsifier.

  2. Question 02

    Take the £1.4m prepay?

    A £1.4m prepay tied to a 99.99% SLA. Watch whether the moderator sides with the most confident perspective.

  3. Question 03

    Out of scope

    A personal legal question. Watch the scope guardrail that runs alongside triage.

Choose one once inside. Each runs from the question to the panel’s recommendation.

Item 3 · How it is built

Under the hood

OpenAI Agents SDK · Python

PatternMulti-agent handoffs with a synthesising moderator
Packagesopenai-agentspydantic
Runnable sourceagents/quorum-board/ entry quorum/board.py
Why this architecture, and what it costsHandoffs model "a different agent should own this now" as a first-class transfer, so each perspective keeps a narrow brief and a clean context instead of one agent role-playing four opinions. The cost is more turns and more tokens than a single call.
Tools and their risk tier
  • state_positionsafeRecord one perspective's position, with a confidence and the key risk.
  • record_decisionsafeRecord the recommendation, the dissenting view, and what would reverse it.

Item 4 · See it work

Watch the panel discuss

Take a look inside