Humanfia

Appearance

About Humanfia ​

We are a small group building the flow around the agent: the method, the referee and the applications that decide whether an agent finishes the work or merely starts it impressively.

The bet ​

A stronger model writes a better function. It does not, on its own, work for eleven hours, notice that the third approach was the right one, keep the constraint it was given in round one, and stop when the job is actually done. That is a property of the flow around the agent, and almost nobody measures it.

Everyone can see that models are getting better. Fewer people are working on the thing that determines whether eleven hours of agent time produces a merged change or a pile of plausible diffs — the loop.

We think that loop is the durable artefact. Models are rented and replaced; a method that works is something you can write down, publish, measure and improve. So we build methods, we give them away, and we score them against each other in public. If one of ours loses, it goes.

That 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 or not at all.

How we work ​

Automation without abdication ​

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

The builder is not the judge ​

The model that wrote the code should not be the only model deciding whether the code is good. Every serious flow we run has a second agent that arrives with no memory of how the work was arrived at and reads the repository instead. This is the single highest-leverage thing we know.

Reward hacking is the default failure ​

Not incompetence. On a long run, the way work fails is that it looks finished: a test narrowed until it passes, a branch special-cased on the input the test uses, a stub with a confident summary on top. Our reviewers are told this is the thing they are most there to catch, and our applications are chosen so that it cannot survive the check.

If it is not measured, it is a preference ​

Anyone can have a theory about how agents should be run. 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, including the parts that need caveats.

Who is behind it ​

Humanfia is maintained by @futrime and built alongside collaborators in academia and industry — the KDA work is joint with MIT HAN Lab, and the loop that became Humanize 1 began at PolyArch.

The code is Apache-2.0. Everything we ship is at github.com/humanfia.

Get in touch ​