Appearance

Open-source agent flows

We build the flow around the agents.

Open-source flows that organise coding agents for long-horizon work, the runtime that runs them, and results checked by outside graders.

Get started View on GitHub
  1. New model.
  2. Same flow.
  3. Better result.

A flow is how agents are organised — who works, who checks, when to stop — rather than a harness tuned to one model, so every new model makes it stronger.

How it fits together

01 · Flows

Flows are open-source Python.

A flow is a directory of Python: which agents take turns, what each is asked, and when it stops. RLAR, Flame Chase, the Ralph loop and the rest are open source, to read, fork or improve on.

Browse the flowverse

02 · Runtime

Humanize runs the flows.

It opens and resumes sessions, keeps a budget in time, cost or tokens, puts the work in a worktree, a container or on another machine, and writes every turn down on one clock.

Meet Humanize

03 · Agents

It drives ten coding-agent CLIs you already use.

claude, codex, dsh, agy, grok, kimi, qwen, pi, opencode and mimo. We hold no API key: the runtime drives the tools you already pay for.

Read the docs

04 · Results

Flows are tested on problems with an outside grader.

Proofs checked by Lean, kernels timed on real hardware, models scored on public benchmarks and competitions. The grader is never ours, so a result does not rest on our own judgement.

See the achievements

05 · FlowBench

FlowBench makes the winning flow the next default.

FlowBench scores flows against each other on long-horizon work, ours included, with the model held fixed. The winner becomes the next default.

About FlowBench

Achievements

Each result links to a dated post that names who produced it and how to check it.

First solves on Lean-Eval v1, the most of anyone30Second by one on total, and 170 of 171 on the archive.MLE-bench medal rate78.2%Two agents taking turns on 75 tasks in six hours. Self-reported.Throughput from one KDA kernel in SGLang+8.7%Qwen3.5-4B at concurrency 1, from one of six KDA pull requests SGLang merged since August.Kaggle top 5%16/39The best tracked result in each of 39 completed competitions; three are final private ranks.Science exams at full marks5/5IPhO, IChO, IOI, IBO and quantum information theory — none graded by an official jury.PutnamBench672/672Every statement proved in Lean 4, joint first on the official leaderboard.MSA indexer, in production6.5×Prefill on B300, bitwise-identical output. Humanfia-reported.ProgramBench3.5%Two models that solve 0% and 0.5% alone, as a builder and a reviewer in a loop.Past the best human kernels1.69×On every MLSys 2026 track: 1.69× on GDN prefill, 1.41× on DSA, 1.17× on MoE.IMO 20266/6Every problem solved by a fully agentic run and machine-checked in Lean 4 — on two different backends.First on SOL-ExecBench L10.7639Ahead of doubleAI's 0.7587, on a track everyone had tuned.First places on SOL-ExecBench7171 of 235 per-kernel leaderboards, the most of any entrant, on 5 October.Files changed to port gem5's build from SCons to CMake567One engineer and a set of agents, running for weeks.

Results to date

≈2
months, June to August 2026
≈10
people

Mathematics and science 12

GPU kernels 8

Humanize

It drives the coding-agent CLIs you already use, through your existing logins, and holds no API key of its own.

claudecodexdshagygrokkimiqwenpiopencodemimo
zsh
$ uv tool install git+https://github.com/humanfia/humanize.git
$ hmz exec -f official/flame_chase

Time, cost and token budgets

Stop on time, cost or tokens, so a run meant to take a night does not take a week.

Every turn on one timeline

Every agent, sub-agent and tool call on a single timeline, opened as a Chrome trace.

Local, container or SSH

Locally, in a git worktree, in a container, or on another machine over SSH.

Read the documentation

News

HOAOct 4, 2026#1 on Lean-Eval v1 first solvesLean-Eval froze its first 128-problem release and moved everything older to an archive. On v1 our GPT-5.6 flow has the most first solves, 30, and is one behind the leader on total; on the archive it has 170 of 171.Read ›HMAOct 3, 202678.2% MLE-bench medal rate in 6 hoursTwo coding agents taking turns over one workspace, each starting fresh, beat both of them working alone on 75 MLE-bench tasks: 78.2% any-medal against 72.4% and 68.0%, in the same six hours. Self-reported, and not on the official leaderboard.Read ›KDAOct 1, 20266 KDA PRs merged into SGLangSince August, SGLang has merged six pull requests carrying kernels our agents wrote, the best adding 8.7% end-to-end throughput. The MLSys contest kernels are public with a benchmark anyone can rerun. One merge elsewhere, Flash-KMeans, was reverted the next day.Read ›HMASep 30, 2026Biohub final: #188 of 3,947, top 5%In August an agent workflow sat 8th on the Biohub cell-tracking public leaderboard. The final private rank is 188th of 3,947, the top 4.76%: lower, but an official finish. Plus the Kaggle tally when only final ranks count.Read ›HOASep 14, 202630/30 IPhO, 6/6 IOI, 100/100 IBOIPhO 2026 theory 30 of 30, IOI 2026 six of six, IBO 2024 theory 100 of 100, IChO 2026 formalized 68 of 68, and quantum information theory 40 of 40. What each score is, who graded it, and what it does not show.Read ›HOAAug 27, 2026672/672 on PutnamBench, joint #1GPT-5.6 in a worker–reviewer loop proved all 672 PutnamBench statements in Lean, at $44.50 a problem. The top of the official leaderboard is now a four-way tie, and what separates the four is cost.Read ›HOAAug 18, 2026#1 on Lean-Eval: 172 proofsThree weeks after second place, a fully agentic GPT-5.6 run is first on the Lean-Eval leaderboard, with 172 research-level problems solved. Every accepted proof is sorry-free and independently re-verified.Read ›HMAAug 15, 2026Two Kaggle finishes in the top 3%Ten agent workflows on Kaggle, with every result sorted by what kind of rank it is. Two closed with official final ranks in the top 3%. Twelve more top-5% results are late-submission estimates, and six live public boards had us in the top 5% on 15 August.Read ›All newsEvery number we have published, newest first — and an RSS feed.Open the news ›

Long-term target

  1. 100Tmodel
  2. 1Ktokens per second
  3. 1Magents
  4. 1Menvironments and containers
  5. 1Mminutes

One flow framework at this scale, working on a problem too large for any single model.

These are targets we are building toward. What we can measure today is on Flow Science.

The code and the results are public.

Every flow is Python anyone can read or fork, and every number on this page links to the post that says how to check it.

Star Humanize on GitHubRead the docs

github.com/humanfia · flowverse · RSS