Debate (MAD)

The debate loop — multi-agent debate, or MAD — puts several agents on the same question and lets them argue. Each round, every debater sees the others’ prior arguments and revises its own; after a set number of rounds, a reducer picks the answer. Adversarial scrutiny surfaces reasoning errors that a single chain would carry to the end, which is why debate shines on contested or reasoning-heavy questions where being challenged improves the answer.

This page explains the loop. For how a loop is chosen, see Coordination loops.

The loop

Rendering diagram…

N debaters answer independently, then exchange peers' prior arguments each round until the final round reduces to one answer

Debate is its own skeleton — a first-class coordinator, not a variant of VMAO:

• Round 0 — the N debater lanes answer independently, with no peer context.
• Round r > 0 — each lane’s prompt carries the other lanes’ round-(r−1) arguments as structured turns, and the lane revises. There is no live agent-to-agent channel: the coordinator arbitrates the exchange by re-injecting peer arguments at each round barrier. Every round is one governed, budget-carved, checkpointed wave, so an operator sees round progress advance on the event log.
• Reduce — after the final round, a reducer collapses the arguments to one answer.

Per-round audit detail rides an extension signal on the event log for observability only; it never feeds back into the loop’s control flow.

Two reducers

• Majority vote — the debaters’ final answers are extracted and the most agreed-upon one wins. This is the Du et al. (2023) form and the shipped default.
• Comparative judge — a judge wave ranks the final arguments and names a winner, with a rotation to control ordering bias and a majority-vote fallback when the verdict can’t be parsed. This is the Liang et al. (2023) form. The coordinator also supports convergence early-stop and a judge-driven early exit that ends the debate once the matter is resolved.

The shipped configuration io.descoped/deep-research-debate@1.0.0 runs a governed N = 3 lanes, R = 2 rounds, majority-vote debate. The comparative judge and the early-exit options are selected through a custom strategy configuration’s parameters.

Running out of rounds, and steering the ones you grant

When the debate hits its round ceiling without resolving, the coordinator does not silently stop — it takes the governed stall path and pauses for a human, exactly as VMAO does at its own cap. You can terminate, accept the partial result, or grant more rounds.

A grant may carry a comment, and that comment reaches the debaters. It is composed into the next round as its own labelled turn — never spliced into the system template — so an operator extending a debate can also say what the extra rounds are for (“neither side has addressed the licensing constraint”) rather than just buying more of the same argument.

The steer is one-shot: the coordinator clears it after the round that consumes it, so a nudge given to break one deadlock does not quietly re-direct every later round.

Knowledge-only vs tool-equipped

The shipped debate configuration is knowledge-only: the lanes reason from the model’s own knowledge and complete with no per-role skill dependency. Binding the research skills to the debater and synthesizer roles at launch turns it into a tool-equipped debate — the lanes gain web tools and the run ends with a synthesized report. Same skeleton, different per-role binding.

Research lineage

Majority-vote debate is Du et al. (2023); the comparative judge is Liang et al. (2023). The compute-matched-baseline framing — debate earns its extra agents only when it beats spending the same compute on a single richer sample — comes from Smit et al. (ICML 2024); per-lane model choice follows the heterogeneous-MAD line. This is why debate sits beside parallel-sample: both spend parallel compute, and the honest question is always which one a given problem rewards.

When to choose it

Choose debate when a question is contested or reasoning-heavy and being challenged improves the answer — analysis with competing interpretations, decisions with trade-offs, claims worth stress-testing. If independent attempts without cross-talk suffice, parallel-sample is cheaper; if the work is decomposable and needs verification, use VMAO.

Where this shows up

• Build: Deep Research ships a debate configuration (io.descoped/deep-research-mad) alongside its VMAO one.
• Reference: the orchestrator’s debate coordinator under Reference.