agentix.consistency

consistency

Self-consistency.

SelfConsistencyModel wraps a model and, on each turn, samples it N times and returns the majority vote — a simple, effective way to damp non-determinism on hard reasoning steps. It's a ModelFn, so it drops into Agent(model=...).

Cost: it makes N model calls per turn. The returned response aggregates the token/USD spend across all N samples, so the agent's budgets and outcome.cost_usd reflect the real cost. Pair with a Limiter to bound the extra concurrency.

Responses are grouped by a key (default: normalized text for a final answer, or the tool-call signature otherwise); the largest group wins, ties go to the first sampled.

SelfConsistencyModel

SelfConsistencyModel(
    model: ModelFn,
    *,
    samples: int = 5,
    key: Callable[[ModelResponse], str] | None = None,
)

Sample the wrapped model N times per turn and return the majority vote.