Five autonomous AI agents, no humans: parallel search finds solutions a single agent can't

A KDD'26 workshop paper posted to arXiv this week ran LLM agents in parallel against a shared score, with no humans in the loop. The testbed was product-to-catalog matching. The author's argument is that the lift in numbers reflects a different shape of search across the population of agents, and the score is the artifact that makes the difference visible.

The paper, Continuous Improvement and Parallel Autonomous Exploration by Andre Rusli, is a 5-page preprint accepted at the KDD'26 Workshop on SciSoc Agents & LLMs. It treats product-to-catalog matching as selective prediction (paper text): an agent either commits a product to a category or abstains, and the reward is qualified coverage on a held-out test set at a precision floor of 95%. The framework wires two pieces together. A leaderboard against held-out data acts as a continuous improvement signal: agents refine against a score that does not leak the test set, and the leaderboard's stability across rounds keeps the refinement cumulative rather than noisy. A parallel substrate lets multiple agents run independent trajectories against that same score, with a moderator agent handling only scheduling and result collection.

A baseline system reaches 33.3% qualified coverage. A single agent on the same task, refining against the leaderboard, hits 47.8-57.4% across categories. Five agents running in parallel reach 62.8-69.4%. The author attributes the gap to two interacting loops: a single agent's improvement over time, and the system's parallel exploration across instances. The first loop plateaus when the same agent runs out of new trajectories. The second loop does not, because the population keeps finding new ones.

The mechanism is the leaderboard. Held-out data gives the agents a reward signal that is consistent across runs without giving the test set away. A single agent refining alone eventually hits a ceiling: the same exploration patterns recur, and the score plateaus. Parallel agents, sharing the score but not the trajectory, explore different regions of the solution space. The moderator is thin on purpose. A heavier moderator would absorb the exploration budget and collapse the parallel substrate back into a single agent with a sidekick. No human reviews solutions; no human curates prompts between rounds.

Read as a benchmark, this is a niche e-commerce bump on a workshop track. Read as a substrate, it is a reusable pattern for how autonomous AI work gets organized: a shared reward on held-out data, a population of agents that explore in parallel, and a thin moderator that stays out of the way. The e-commerce testbed is the proof that the pattern runs end-to-end without humans in the loop. The pattern looks more like research infrastructure than a demo, and the metric is the part that does the coordination.

The paper is a single preprint, not yet peer-reviewed in the published-proceedings sense, and the testbed is one narrow task on one vendor's product catalog. The qualitative-difference claim is the author's interpretation of the runs, not an independent measurement. A falsifier would be a parallel substrate failing to break a single-agent ceiling on a different solution space; the paper does not run that experiment. Outside the testbed, the substrate's behavior on non-retrieval tasks is not addressed, and the "qualitatively different" claim is grounded in the author's reading of the trajectories, not a formal diversity metric.

The framework offers a coordination surface: a leaderboard on held-out data that lets a population of agents self-improve without leaking the answer, and a moderator thin enough to stay out of the way. That pattern has not been tested past product-to-catalog matching. The next move is replication on a different solution space, ideally with a second team and a second dataset.