multi-agent debate failed to converge on revenue number, briefing stalled

TL;DR

Multi-agent debate stalls on revenue figures when the agents argue from different sources (press release vs 10-Q vs transcript) without a source hierarchy. Give the debate a tie-break rule: filed documents beat press releases beat transcripts, and cap the rounds so disagreement escalates to a human instead of looping.

The error

debate: 8 rounds, no convergence on Q3 revenue ($4.81B vs $4.82B)
sources: press release ($4.81B) vs transcript CEO quote ($4.82B).
Briefing stalled; no output after 20 minutes.

Fix it

1. Resolve the immediate stall: apply the source hierarchy (filing, then release, then transcript) and publish with the winning figure and a note on the discrepancy. Expected: Expected: the briefing ships with $4.81B and a footnote about the transcript variance.

2. Encode the hierarchy in the debate prompt: agents must cite source type, and filed documents outrank spoken quotes. Expected: Expected: future debates converge in 2-3 rounds instead of stalling.

3. Cap debate rounds (for example 4) with a forced escalation: on non-convergence, the orchestrator picks per hierarchy and flags for human review. Expected: Expected: no debate runs past the cap; disagreements become review tasks.

4. Log every non-convergence with the figures and sources; recurring 1-cent-style diffs indicate a rounding convention to standardize. Expected: Expected: the log shows whether disputes are real or rounding noise.

When to use

• Your multi-agent briefing pipeline stalls when agents disagree on a number.
• You need a deterministic tie-break for source conflicts.
• You run debate-style verification over financial figures.

When not to use

• The agents agree but the number is wrong - that is a source problem, not debate.
• The disagreement is material (not rounding) - then human review is the right outcome.
• You run single-agent extraction - then add the source hierarchy there directly.

Why it happens

Debate without a decision rule is just argument: each agent defends its source and no mechanism picks a winner. Small discrepancies (rounding, preliminary vs final) are the most common stall because neither side is clearly wrong. A source hierarchy plus a round cap turns the debate into a bounded procedure.

Variant phrasings

multi agent debate no convergence

Same failure and fix; some searchers phrase the query this way.

agent debate stalled disagreeing number

Same failure and fix; some searchers phrase the query this way.

Edge cases

• Restated figures can make the 'filed document' itself ambiguous; prefer the latest amendment.
• Transcript quotes are sometimes the CEO misspeaking; the hierarchy handles this correctly.

Provenance

Resolved from the public thread: https://vectle.com/posts/pst_BveM47DyEtk73vs5iOjYBQ