Roundtable

Responding to a request for proposal is a team sport. Someone researches the client, someone scopes the work, someone estimates, someone prices, someone checks the contract terms, someone writes, and someone tears the draft apart before it goes out. It takes a week and the best people in the company, every time.

Single-agent systems are good at tasks. They are bad at judgment calls that need more than one perspective: is this scope realistic, is this price defensible, is that clause a risk, does this draft actually answer what the client asked. Real teams solve that with disagreement. Most multi-agent demos solve it with a supervisor that tells workers what to do and never lets them argue.

You give Roundtable the RFP and a connection to the firm's knowledge base: past proposals with outcomes, case studies, the rate card, the margin policy, the legal playbook, the style guide. A Chair opens the meeting. The Scoper turns the RFP into a requirements matrix with a page citation on every line. The Researcher brings client context and comparable past deals. The Estimator spawns one sub-estimator per work package and reconciles them, explaining any disagreement. The Pricer prices from the rate card and policy and cites a past deal for every deviation. Risk flags contract terms against the playbook. The Writer drafts sections in parallel, every claim cited.

Then the Critic reads the draft the way the client's evaluation committee will, scores it against the RFP's own criteria, and objects, with evidence. Owners defend or revise; the Arbiter settles what they cannot; anything with even evidence goes to you, both sides on one screen. The output is the proposal, a pricing sheet, a risk register, a client Q&A list, and the minutes: who said what, what changed, why.

The demo is a fictional 3PL's 34-page RFP answered by a fictional 25-person consultancy, including a 40% estimate disagreement on one package, a price objection argued to a ruling, and an unlimited-liability clause escalated to the human seat.

Three primitives. A blackboard: one versioned store of typed entries, each with an owner and evidence; agents read anything and write only what they own, so two agents can never edit the same thing at once. A message bus: typed messages, always routed through the Chair's turn schedule, never ad hoc; the transcript is the minutes. A protocol: a fixed sequence of rounds the Chair runs, with turn budgets, round limits and objection limits, so nothing loops forever.

Objections are entries, not chat. One needs a target, a claim, a citation and a proposed resolution, or the Chair rejects it before anyone spends a turn. The owner defends with citations or revises; after three rounds the Arbiter rules with a written reason; after five arbitrations in a meeting, a person settles the rest. Fan-out is how the heavy lifting gets done: one sub-estimator per package, one writer per section, each with a narrow context pack and a small budget, reconciled by the parent in an entry the Critic can object to.

The human is a participant. From the meeting room you can answer an open question, override an entry, veto a clause or settle a dispute. The Chair treats that as the highest-priority turn and resumes from the affected round, and the minutes record it.

Agents disagree on the record. Structure keeps the argument productive instead of endless. Every number in the proposal traces to the rate card, a past deal or the requirements matrix, and every claim to a case study; the Critic's last pass rejects anything that does not, and the Chair will not finalise until it passes.

It is evaluated like a team, not a chatbot: requirement recall, estimate consistency, policy compliance, proposal quality against human-written baselines, objection validity and resolution correctness, protocol conformance on recorded meetings. A change that lowers a score does not merge. Claude Opus 5 sits in the Chair, Critic and Arbiter seats; Sonnet 5 handles research and drafting volume; a vLLM path exists for firms that keep documents in-house.

Started September 2026. Phase 0 of 5. Nothing runs yet; what exists is the collaboration protocol, the agent specifications, the architecture, the evaluation strategy and the demo scenario, published before the first line of code so the build can be judged against what was promised. Each phase has a definition of done and a GitHub issue.

It is the sister project of Dispatch. Dispatch is about running a back office; Roundtable is about agents deliberating. They share the same principles: cite everything, evaluate everything, keep a person at the table.

Every multi-agent system I have shipped taught me the same thing: the hard part is not getting agents to talk, it is getting them to stop, to converge, and to leave a record a human can audit. Roundtable is my answer to that, in the open, on a problem every services business recognises.