Skip to content

ProjectIn development · not released

FlowBench It holds the model fixed and scores the flow

For agent tasks that take hours. Existing tests and benchmarks score every run.

FlowBench · boardNot released
#FlowModelCheckScore
1ralph_loopsamereal--.-
2rlarsamereal--.-
3flame_chasesamereal--.-
Example flows. No scores published yet.
Task length
Hours
Of agent time, per task.
Variable
1
The flow around the model.
LLM graders
0
Every score comes from an existing check.
Numbers published
—
None yet.

01The gap

Benchmarks compare models, but rarely the flow around them ​

Read the same run as
  • Modelthe variable
  • Loop around itone fixed harness, or one turn
  • Task lengthminutes
  • Scoreoften graded by a model

Result: a ranking of models

On tasks that take hours, the loop around the model can matter more than the model. FlowBench measures that effect.

02Methodology

Everything is held fixed except the flow ​

Each task runs under every flow, and the same existing check scores every run.

Resulta ranking of flows

Rule 1: Tasks take hours of agent time

Single-turn tasks do not test the flow around the model.

Rule 2: Scores come from checks, never from a language model

Every task uses a check that already existed: the code runs faster, compiles or passes its tests.

Rule 3: Model, tools and budget never change between runs

Rule 4: Flows are scored exactly as users run them

Flows run unchanged in Humanize, the runtime users install.

03Evidence so far

Three earlier results where the loop moved the score ​

Each comes from one of our other projects, with the model unchanged. These are not FlowBench results.

PutnamBench · 50 problems Same: Kimi-K3
Its own coding CLI4 / 50
Inside a Humanize flow47 / 50
IChO 2026 · 68 subquestions Same: GPT-5.6 Sol
A plain /goal32 / 68
With formalization and proof review68 / 68
MLE-bench · any medal Same: Opus 5 and GPT-5.6-sol, taking turns
GPT-5.6-sol goes first73.3%
Opus 5 goes first78.2%

05Status

Posts about FlowBench so far ​

The release will be announced in the news and at github.com/humanfia.

  1. Jul 8, 2026blogAPI · CLI · flowModel, Tool, Flow: 2, 18, 50 of 50
  2. Aug 17, 2026blogArchitectureHumanfia: From a Review Loop to Four Layers and a Referee
  3. Sep 27, 2026blog4 → 47KDA²: Kernel Design Agents (KDA) Optimize Kimi Delta Attention (KDA)
  4. NextNot yetFlowBench release and first cross-flow numbersThe first numbers will cover the flows on this site, including any where ours lose.