← Back to all posts

Harnessing Agentic AI Systems: Voting / Consensual Ensemble Pattern

Problem 15 of 15: producing a verdict no single model can fake. The Voting / Consensual Ensemble pattern, the Committee Paradox anti-pattern, the Generator–Evaluator Loop and Live-Environment Evaluators frontiers — one table, a short discussion, the key insight, and the important references.

harnesspattern-languageagentic-aiseriesverificationensemblesevaluation

Problem 15 of 15 in the Harnessing Agentic AI Systems series — read the index for the framing. Previous: Sequential Pipeline Routing Pattern.

The Problem — Producing a verdict no single model can fake

One judgment is unreliable, and self-evaluation is reliably lenient. The system must decide what is good without depending on a single model's opinion — and the frontier has evolved the gate into an iterative critic (F1) with hands (F5).

Field P16 — Voting / Consensual Ensemble (pattern) A10 — The Committee Paradox (anti-pattern) F1 — Generator–Evaluator Loop (frontier) F5 — Live-Environment Evaluators (frontier)
Forces / Smell Reliability vs cost; independence vs one family; agreement vs ground truth. Agents reviewing each other; no exit condition; "making progress" without converging. Independent feedback vs self-review; iteration vs cost. Live interaction vs static scoring; reality vs CI speed.
Solution / Anti-solution Query multiple independent model setups with identical prompts; use harness code to calculate majority agreement. "More debate equals better decisions." Separate the generator from the evaluator — the critique is the next iteration's input. Give the evaluator hands — the Playwright MCP against the live page, "the way a user would."
Consequences / Failure A statistical signal where verifiers are king; cross-review replaces self-review. An infinite loop with a nicer name; tokens burn, no verdict arrives. A skeptical standalone evaluator is tunable where self-criticism is not; the full-stack run beat the solo run. The verifier uses the artifact instead of reading it; only a verifier with hands finds the broken wiring.
Tradeoffs / Refactoring Cost scales linearly; correlated members vote as one and add nothing; agreement measures preference, not correctness. Termination as a system property — threshold, budget, breakpoint; P16's aggregation as the disagreement rule. "Over 20x more expensive" — worth it only when output quality justifies the bill. Wall-clock: runs stretched to four hours; reserved for the final slow-loop gate.
Evidence LMArena (leaderboard); mob self-review — 79% of 25,264 agent PRs (mob programming remastered). AutoGen's termination as a first-class design concern (paper). Anthropic's harness (post). Anthropic's harness (post).
Related Composes with P13; aggregation answer to A10; statistical cousin of F1. Is the absence of P13 and P16's termination; orchestration form of A2. Composes with P16; pairs with F2. Slow-loop complement of P11; composes with F1.

Discussion

The verdict problem is where the "system, not the agent" framing is most visible: self-evaluation is reliably lenient — "agents tend to respond by confidently praising the work" — so the system must not depend on a single model's opinion. The pattern makes verification statistical: independent judgments aggregated by code, and the honest limit is that agreement measures preference, not correctness. The frontier evolves the gate into a critic with agency (F1) and hands (F5), both at an honest price; the committee without an aggregation rule is an infinite loop with a nicer name.

Key Insight

Agreement is a signal, not a ground truth. Independence is the whole game — correlated members vote as one and add nothing — and termination is the difference between a debate and a decision. The ensemble replaces self-review with cross-review, and it is never a ground truth.

References

LMArena (lmarena.ai); AutoGen (arXiv:2308.08155); Anthropic, Harness design for long-running application development (post); archive: verifiers-are-king, mob programming remastered.