Humanfia

Appearance

About Humanfia ​

We build the flow around the agent: the method, the referee, and the applications that show whether an agent finishes the work or just starts it impressively.

The bet ​

A better model writes a better function. It does not, on its own, work for eleven hours, notice that the third approach was the right one, hold on to a constraint from round one, and stop when the job is done. That is the loop around the agent — the thing that decides whether eleven hours of agent time ends in a merged change or a pile of plausible diffs — and almost nobody measures it.

We think the loop is what lasts. Models are rented and replaced; a method can be written down, published, measured and improved. So we build methods, give them away, and score them against each other in public. When one of ours loses, it goes.

It is also why we do not build a model, resell an API or hold anybody's key. The runtime drives the coding-agent CLI you already log into. If we are adding value, it is above that line.

How we work ​

The person decides what done means ​

People own the intent, the tradeoffs and the definition of done. Agents are leverage inside that, not a replacement for it. In Humanize 1, the person is one of the agents in the loop.

The builder is not the judge ​

The model that wrote the code should not be the only one deciding whether it is any good. Every serious flow we run has a second agent that arrives with no memory of the work and reads the repository instead. It is the most useful thing we know.

Reward hacking is the normal failure ​

Not incompetence. On a long run, work fails by looking finished: a test narrowed until it passes, a branch special-cased on the input that test uses, a stub with a confident summary on top. Our reviewers are told to look for exactly that, and our applications are picked so it cannot survive the check.

If it is not measured, it is a preference ​

Anyone can have a theory about how to run agents. FlowBench is us agreeing in advance to find out whether ours are right, and to delete the ones that are not.

Build in public ​

The runtime, the flows and the reproduction code are open source. When we publish a number, we publish what it takes to check it, caveats included.

Who is behind it ​

Sihao Liu proposed Humanize 1 at UCLA PolyArch, and Ligeng Zhu built the team around it. Everyone else is on the team page. Each result is credited at the top of the post that reports it rather than folded into a team byline.

The code is Apache-2.0, at github.com/humanfia.

Get in touch ​