AI Agent Operations

When two agents disagree on the fix,
a structured debate beats a coin flip

Agent Debate is what happens before Consensus Arbiter. When two agents propose conflicting actions, debate runs a 1-to-3-round structured exchange where each side defends its proposal and critiques the other. The arbiter picks the winner from the post-debate proposals, which usually look different (and better) than the pre-debate ones.

Get Started Talk to Sales
app.novaaiops.com / agent-debate
● LIVE
1-3
Rounds, configurable
Visible
transcripts in bundles
Auto
on conflict
Time-boxed
60s default
Round Structure

Propose, critique, refine

Round 1: each agent proposes its action. Round 2: each agent critiques the other's proposal and refines its own. Round 3 (optional): final position. The number of rounds is configurable per conflict severity, quick disagreements get one round, ambiguous ones get three. The transcript is captured in the Decision Bundle for both bundle entries.

  • ✓
    Round 1 · propose: each agent presents its proposal with evidence and confidence
  • ✓
    Round 2 · critique + refine: each agent reads the other's proposal and updates its own (or holds firm)
  • ✓
    Round 3 · final (optional): when the gap remains wide, a final round forces explicit positions
app.novaaiops.com / agent-debate · structure
Time Box

Debate is bounded, not endless

Debate is time-boxed (default 60s total). If the rounds run out without convergence, the original proposals go to Consensus Arbiter unchanged. Time-boxing prevents debate from becoming a delay tactic when the underlying incident is on the clock.

  • ✓
    60s default total: short enough for incident response; tunable per class up to 5 minutes
  • ✓
    Falls through to arbiter: no convergence means the arbiter still decides, debate never blocks the response
  • ✓
    Configurable per class: low-risk classes get tighter boxes; security-relevant ones get longer
app.novaaiops.com / agent-debate · timing
Transcript Visibility

Both sides recorded for postmortem

Debate transcripts land in both agents' Decision Bundles. When a postmortem reviewer asks "why did we pick this fix?", the transcript shows the alternative considered, the critique that landed, and the final position. The decision is auditable, not magical.

  • ✓
    Bundled to both agents: transcript appears in postgres-doctor's and cache-warmer's bundles for the same incident
  • ✓
    Read-only: transcripts are never edited after the debate ends; they are the historical record
  • ✓
    Searchable: NovaQL can query transcripts by phrase or position so audit can find specific arguments
app.novaaiops.com / agent-debate · transcript
Convergence Patterns

When debate works and when it does not

The page tracks convergence rate: percent of debates where both agents agreed by the final round. Convergence is good; persistent non-convergence is a tuning signal, usually one agent has stale prompts or one has access to context the other lacks. Patterns of non-convergence get flagged for prompt review.

  • ✓
    Convergence per agent pair: high convergence pairs work well together; low convergence pairs may need scope adjustment
  • ✓
    Pattern detection: when a pair persistently disagrees on a class, the page flags the class for prompt review
  • ✓
    Prompt-fix loop: fixing the prompt that drives the disagreement raises convergence and lowers arbiter load
app.novaaiops.com / agent-debate · patterns
Video walkthrough coming soon

Subscribe to Nova AI Ops on YouTube for demos, tutorials, and feature deep-dives.

Disagreement made productive

Debate is the cheapest way to get a multi-agent system to look like a careful team rather than two confident strangers.

Get Started Request a Demo