We build the flow around the agents: the flows, the runtime that runs them, the benchmark that compares them, and the projects that test them on hard problems. These are the people who do it.
We think the method around the agent outlasts the model
A better model writes a better function; it does not, on its own, work for eleven hours, hold on to a constraint from round one, and stop when the job is done. That is the loop around the agent, and we think the loop is what lasts. Models are replaced every few months; a method can be written down, measured and improved, and it gets better with each new model. So we open-source our flows, benchmark them against other flows, and drop ours when they lose.
How we work
- The person decides what done means. People own the intent, the tradeoffs and the definition of done. Agents do the work within those decisions.
- A separate agent reviews the work. Every serious flow has a second agent that arrives with no memory of the work and reads the repository instead.
- Reward hacking is the normal failure. Long runs fail by looking finished — a narrowed test, a special case, a stub under a confident summary. Our reviewers hunt for exactly that.
- We only claim what we have measured. FlowBench tests our flows against others, and we remove the ones that lose.
- Open source. The runtime, the flows and the reproduction code are open source. Every published number comes with what is needed to check it.
Credit
Each result is credited at the top of the post that reports it, not folded into a team byline. Everything has been built with many more people than are named above — NVIDIA Research, MIT HAN Lab, UCLA PolyArch, Tsinghua, Nunchux, and a long tail of community contributors. The full list is in each repository's contributor graph: Humanize · the flowverse · this site.
Get in touch