Back to explorer

The agency benchmark

The explorer measures a tool in isolation. This measures tools and skills as workers inside a real project. Every run is a ticket for one fictional client — Verdant Labs, maker of the Canopy smart hydroponic garden — that fans out into typed, chained deliverables: a plan feeds an architecture diagram, which feeds a build, which feeds the invoice. Each worker sees its teammates' real output; the whole delivery is judged as one body of work.

52 live chains 12 A/B propagation pairs 3 driver stacks methodology 39864b312e7767ef

Does a skill's value carry downstream?

Two arms of the identical ticket, differing by one worker — a skill-taught vs an untaught (@bare) plan or build. Everything after the pivot measures whether that one change propagates. A positive delta means the skill made the rest of the project better, not just its own stage.

build pivot +0.01

mean chain delta · n=3 pairs · 1 skill-helped, 1 skill-hurt

plan pivot +0.10

mean chain delta · n=9 pairs · 6 skill-helped, 1 skill-hurt

The plan pivot lifts the whole chain (a better PRD compounds — twice it rescued a later stage from 0 to a full pass); the build pivot is near-flat (a better builder mostly just builds). Directional at this n, with run-to-run variance visible in the tail.

pivotticketskill arm vs barechain Δdownstream stage deltas
buildlaunchskill vs bare0.00billing 0.00
planlaunchskill vs bare+0.05chart 0.00build 0.00billing 0.00
planexpansionskill vs bare+0.38creative +1.00billing 0.00
buildexpansionskill vs bare-0.20billing 0.00
buildlaunchskill vs bare+0.23billing 0.00
planquoteskill vs bare0.00billing 0.00
planlaunchskill vs bare+0.53build +1.00billing 0.00
planexpansionskill vs bare+0.03build 0.00billing -0.75
planquoteskill vs bare0.00billing 0.00
planlaunchskill vs bare+0.23chart 0.00build 0.00billing 0.00
planquoteskill vs bare-0.38billing -0.75
planexpansionskill vs bare+0.07creative 0.00build 0.00billing 0.00

Driver stacks across the same work

Same tickets, same stage workers (rotation held constant) — only the driving model + harness changes. This measures the stack, not the model in isolation: PI-over-OpenRouter for glm and MiMo, Claude Code over ACP for Sonnet.

mimo-v2.5
75%
mean chain score · n=13
glm-5.2
71%
mean chain score · n=28
claude-sonnet-5
70%
mean chain score · n=11
templatemimo-v2.5glm-5.2claude-sonnet-5
audit77%n267%n580%n2
expansion74%n267%n570%n2
incident95%n278%n593%n1
launch81%n375%n390%n2
quote87%n269%n563%n2
status36%n272%n536%n2

Note the status row: both frontier stacks drop sharply on short comms tickets where glm holds steady — a harness-fit effect (heavy harnesses over-engineer trivial work), not a model-quality ranking.

Chains

Every run is a real deliverable chain. Click one to see its stages, workers, and the deterministic coherence checks against the world-file.

How this is scored. Each stage's artifact is graded by the same deterministic checkers as the explorer. Then the bundle gets a deterministic coherence check against the world-file (brand palette propagated, requirement ids traced, invoice math correct, figures canonical), and a frontier judge panel (Opus 4.8 + Gemini, never a run's own driver) reviews the whole delivery. Chain score = mean stage checker × (0.5 + 0.5 × coherence). Full rules on the methodology page.