Saphan

Research — study AI agents at work, not in a benchmark.

Saphan Studio records every step a fleet of AI coding agents takes on real software: every task, every run, every independent review and every human decision. For researchers, that is a dataset and an instrument in one.

What you can study

Real work, fully recorded

Each agent run with its model, time, cost and outcome; each review with its verdict and findings; each human decision with its reason and time. Nothing to scrape or clean: the record is the data.

The effect of the rules on the work

Change one rule — who reviews, how far an agent may act alone, when a person must approve — and measure what happens to quality (work sent back by reviewers, rounds per change) and to speed (time from task to accepted change, idle time). Every rule change is dated, so each one can be studied as a before and after.

Models side by side, under the same conditions

Several AI vendors on the same tasks, each checked by a reviewer from a different company, with every disagreement recorded.

Reproducible by design

Decisions are signed, the record is only ever added to, and every measurement comes with the command that produced it.

First findings, from our own fleet

Our earlier work on running agent fleets, built on our own methodology and the Saphan protocol, has already led to filed patent applications. Now we are giving it to the public: Proven, an open-source project built end to end by one person directing a fleet of agents, and Saphan Studio as a tool for researchers that records all of it and lets you change the rules a fleet works under. The Open Agent Hackathon, October 2026, is where we show it in public. The findings below come from that week.

Volume with one owner

In two days, across two repositories: 819 commits and 140 changes accepted by the owner, each one reviewed before it was accepted.

The second reviewer matters

In a security review of the agents’ own code, two reviewers from different AI companies disagreed on two of the first three parts: one passed the code, the other rejected it and proved the problem with failing tests. Both problems were fixed the same day.

The bottleneck was approval, not agents

Work ran at full speed while the owner was at the keyboard and nearly stopped overnight, with agents and machines idle, waiting for a human signature. That finding led to a mode in which the owner approves a batch of work before leaving and the fleet continues on its own, with every irreversible step still waiting for a person. We are measuring its effect now.

We hold months of such records, and a log of every rule change with its date.

Academic licence

Saphan Studio Team, free for research and teaching

For non-commercial research and teaching at universities and research institutes: export of runs, reviews and decisions as a dataset; an experiment mode, where you define variants of the rules and every run is tagged with the variant it ran under; and help from the team with your first study.

One thing we ask

Share what you learn. Publish it, and tell us, so the next fleet works better.

More of Saphan Studio

Governance — decisions on the record →Security — untrusted agents, provable runs →Enterprise — your machines, your record →Protocol — how the fleet works →

Apply for an academic licence — tell us your institution, your question and the size of the study