Proofreaders

Pitstop: a supervisor agent that learns to catch drift in agent swarms Long-running agent teams fail quietly. A sub-agent drifts off task, hallucinates a number, or follows a bad instruction, and nobody notices until the final answer is wrong. That makes them impossible to walk away from, especially in finance, where one bad figure invalidates the whole analysis. Pitstop is an optimisation layer for agent supervision. It uses GEPA (Genetic-Pareto prompt optimisation) to make the supervisor recursively better at its job: it reads agent traces, identifies drift, and evolves its own prompts by comparing candidates against the best-performing prompts on its Pareto frontier. As a proof of concept, we built a swarm of financial research agents that use You.com's financial research API to generate reports. We generated synthetic data to bootstrap the GEPA loop, optimised the supervisor, and tested it against a baseline supervisor overseeing the swarm. Traces from each run feed back into the GEPA loop, so the supervisor keeps improving as the agents work. Results on drift detection: Nemotron Nano: 82% → 88% accuracy after optimisation, outperforming the unoptimised Nemotron Super model (86%) at a fraction of the cost. Nemotron Super: 86% → 96% accuracy after optimisation.