Sakana Just Found the Point Where Adding More AI Research Agents Stops Helping

The number is 100. Scale Sakana AI's open-ended discovery agent from 10 copies to 100 running in parallel, and it matches the human novelty baseline exactly: 0.089 to 0.089. Push it another 10x, to 1,000, and the score stalls at 0.088. Worse, 10-20% of everything the bigger swarm produces degenerates into unusable noise, right at the scale the field is now racing toward.

That result landed the same week as two more data points that, read together, undercut a quieter assumption running through the AI-for-science field: that if a research agent works, running more of them in parallel gets you more research. It doesn't, at least not for free, and one benchmark suggests that without a filter, more agents can mean more confidently wrong output, not more discovery.

The dial: 10, 100, 1,000

Sakana AI's Picbreeder experiment, published July 11 alongside the underlying paper "In Search of the Ingredients of Open-Endedness" (arXiv 2605.23908, GECCO 2026), replicates a specific piece of AI history: Picbreeder, the website where humans spent years collaboratively evolving a library of images from small neural networks with no target objective — the canonical example of open-ended, human-driven discovery. Sakana swapped the humans for frontier vision-language models (Gemini 2.5 Pro came out on top, ahead of Gemini 3 Pro Preview and two Qwen3-VL variants) and measured "semantic recall": for each of the 1,824 object classes in the THINGS dataset, the closest match anywhere in the agents' evolved archive, summed across all classes.

Run at a default setting, the best VLM configuration scored 0.087 against a historical human baseline of 0.089. Then the team ran the scaling experiment: 10 agents sharing an archive, then 100, then 1,000, each condition repeated across 6 seeds and 2,000 sessions. At 100 agents, the score hit 0.089 — matching the human number exactly. That's the headline most coverage missed: parallel VLM agents can, at the right scale, replicate a human open-ended-discovery benchmark.

What happens when you turn it past the sweet spot

Ten times more agents than the condition that matched humans did not produce a better score — it produced a statistically flat one. And the paper reports a second cost that a single summary number hides entirely: at 1,000 agents, the shared archive is, in the authors' words, "rife with high-frequency, uninterpretable…psychedelic patterns" making up 10–20% of the total output. Scaling the agent count past its sweet spot didn't just stop helping — it started actively degrading a chunk of what the system produced, and the aggregate recall score wasn't sensitive enough to show it. If you only checked the headline metric, you'd never see the archive was getting worse in places.

Meanwhile: mass-producing papers got easy

FARS (arXiv 2606.31651, submitted June 30 by Analemma AI) is a fully automated research system — ideation through manuscript — deployed at real scale rather than benchmarked on a handful of curated tasks. Its live deployment ran for 417 hours on a 160-GPU cluster, consumed 21.6 billion model tokens, and cost roughly $186,000 in aggregate, producing 166 complete papers across 67 fine-grained AI/ML subtopics — an average of about 2.5 hours, 130 million tokens, and $1,120 per paper. Throughput at that price point, on that timeline, was not possible a year ago.

The same paper is honest about what that throughput bought. FARS collected 282 structured reviews from volunteer reviewers across 140 of the 166 papers, and reports the results were "review-worthy and occasionally strong" but exposed "recurring failure modes in narrow experimental scope, methodological limitations, and integrity issues." Scaling the pipeline scaled the paper count. It did not, on its own, scale the fraction of papers a reviewer would trust.

The part nobody scaled: judgment

PseudoBench(arXiv 2606.18060, June 16) is the sharpest version of the same point. It's a 200-claim adversarial benchmark, spanning five domains, built specifically to test whether an autonomous research agent can recognize and refuse a pseudoscientific premise instead of dutifully building an experiment-to-manuscript pipeline around it. Run end-to-end against seven state-of-the-art auto-research agents, the benchmark found near-zero refusal rates across the board — the single best-performing agent still only resisted 27.4% of the adversarial claims. Worse, the paper notes that stronger agents tended to package the pseudoscience in more sophisticated scientific language, making the output more persuasive, not less.

Put next to FARS, that's the uncomfortable version of the throughput story: the capability that scales cleanly with more compute and more parallel agents is the ability to produce a fluent, well-formatted research artifact. The capability that doesn't scale for free — that in fact needs to be deliberately built in, not assumed — is knowing whether that artifact deserves to exist.

What this means for reproducible science

None of these three results argue against running agents in parallel, or against automating the mechanical parts of the research pipeline — FARS's 166 papers and Sakana's 100-agent match with the human baseline are both genuine capability demonstrations. The argument is narrower and more specific: agent count and compute budget are not, by themselves, a proxy for research quality, and past a certain scale they can actively work against it — more noise in Sakana's archive, more confidently-wrong manuscripts in PseudoBench's adversarial test. The variable that has to be scaled deliberately, separately from parallelism, is verification: something that checks each output before it counts as a result, rather than trusting that more attempts converges on more truth.

That's the same conclusion this newsletter keeps arriving at from different angles — model size wasn't the bottleneck two weeks ago, and raw parallel throughput isn't the answer this week either. The thing that scales trust isn't more agents running longer. It's an independent check on what they hand back.