NeuroAI NEUROAINEUROAI.SITE
China AI, reported clearly
1,300+ verified stories on models, robotics, BCI and the business of AI.
Browse stories
ESC

32 Agents, 18.9x Faster. The Gain Is in the Split.

Thirty-two agents worked on one coding job. The reported speedup was 18.9x.

2026-11-22 · 3869 words · NeuroAI
32 Agents, 18.9x Faster. The Gain Is in the Split.

Thirty-two agents worked on one coding job. The reported speedup was 18.9x. If the work had divided perfectly, the number would have been 32.

That gap is the story.

The missing 13.1

Subtract. 32 minus 18.9 leaves 13.1. That is not slack, and it is not a lazy agent. It is the price of coordination, paid every time one agent hands state to another.

Thirty-two agents can form 496 pairs: 32 times 31, divided by 2. Not every pair talks. But every conversation that does happen costs a message, and a message produces no code.

Replit published the study. It covers coding, and the published detail is one number and one name.

0.59 per agent

Divide 18.9 by 32 and you get about 0.59.

Read that as an average, not a marginal rate. The first few agents take large, separable chunks and add a lot. The last few take slivers and add very little. Sum the contributions and the mean lands at 0.59.

The average hides the ordering, and the ordering is what you need to decide where to stop adding agents. It describes the shape of the past, not the next hire.

Where the ceiling sits

Take the parallel-speedup formula: speedup equals 1 divided by the serial share plus the parallel share divided by 32. Set the result to 18.9 and solve. The serial share comes out near 2.2%.

That is our arithmetic, not Replit's. But it fits the shape of a parallel system, and it has a consequence. If 2.2% of the work cannot be split, no number of agents gets past 1 divided by 0.022, or about 45x.

At 64 agents the same curve tops out near 27x. Doubling the agents adds roughly 8x. After that, each addition buys less than the one before.

The precision is only as good as the input. 18.9 may be rounded, or it may be one run. Treat 2.2% as a direction, not a constant.

The strongest single model was already in the room

The same study names GPT-6.2 Astra as the strongest single model.

That sentence changes how 18.9 reads. If multi-agent work were a substitute for weak models, the result would be quantity covering for quality: one model is not enough, so run 32. But the best single model was available, and 32 agents still produced 18.9x on top of it.

So the gain is not a model upgrade in disguise. It sits in how the work was organized.

One hole remains. The report does not say whether all 32 agents ran the same model. If they did, the 18.9 belongs entirely to splitting and handoff. If they did not, part of it belongs to model mix. The two cannot be separated from what was published, and the report does not try.

The handoff is the tax

What passes between agents is not code. It is state: where I got to, what I changed, what I found.

That state has no standard format. Written long, the receiver has to read before acting. Written short, the receiver has to guess. Each of the 32 agents decides for itself, so there are 32 dialects.

Errors surface at the third hand, after the misunderstanding has been copied into work that other agents have already built on. In parallel, a mistake is not just wrong. It is wrong at the scale of the whole fleet.

This is why the per-agent contribution falls. Early agents inherit whole blocks and hand off rarely. Later agents inherit fragments and have to re-explain context at every boundary. Explanation yields alignment, not output.

Seams decide the multiple

You cannot give a rock to 32 people and tell them to strike it at once. The next blow depends on where the last one landed. You can do that with code, because code already arrives divided by functions, files and tests.

The seams are a property of the task, not of the agents. The agents' job is to find them and push 32 units of effort through. Finding them takes time that cannot be parallelized: someone has to read the job and try it before it can be cut.

Change the task to one with fewer seams and the same 32 agents will report a smaller multiple. The multiple is not an attribute of the model. It is a joint property of the task's shape and the organization imposed on it.

Three questions before the next multiple

The study gives 32, 18.9 and one model name. It does not give the task list, the failure count or the cost.

That is enough for one conclusion. Thirty-two units of effort did not become 32 units of output. The missing 13.1 went into splitting and handoff, and the ceiling is visible even though the bill is not.

The gain is real. It is faster because the work was cut, not because the model was smarter. The next time a multi-agent multiple is quoted, ask three things: how many agents, which single model was the baseline, and what the task was.

Without those three, the number is a slogan.

*Disclaimer: This article is based on publicly available information and is for reference only. It does not constitute investment advice.*

Related coverage

Cite this article

NeuroAI (脑机网), "32 Agents, 18.9x Faster. The Gain Is in the Split," 22 November 2026, https://neuroai.site/a/na-model-replit-32-agents-1122

Machine-readable: BibTeX · RIS · JSON · Markdown

More in “Foundation Models” → · Back to home · Markdown version

China AI, reported clearly
1,300+ verified stories on models, robotics, BCI and the business of AI.
Browse stories