Learn

Benchmarks

What one real index looks like when you grade every session: 2,960 sessions on one laptop, 34% cleared the evidence line. The numbers the landing page used to carry, kept here.

All sections

These are measurements from one machine, not a leaderboard. They answer one question: when you grade every session an agent ran, how far did each one actually get?

One laptop, 2,960 sessions, graded

Source: archie stats on the author's laptop, 2026-09-02, over an index of 2,960 sessions. Every figure is copied from that output; the six rows are the whole index and the percentages sum to 100.

Rung What the session left behind Sessions Share
0 Nothing verified, or still running 1,490 50.3%
1 The agent said it was done 7 0.2%
2 Some files on disk changed 449 15.2%
3 A test or a build passed 808 27.3%
4 A commit landed in git 120 4.1%
5 CI or a deploy went green 86 2.9%

Evidence starts at rung 3. Half of the sessions never got far enough to tell, and rungs 1 and 2 are things an agent can produce without doing much. Only the last three rows left a trace someone else can check: 1,014 sessions, 34.3%, cleared that line.

What the rungs mean, and why rung 3 needs a captured exit code, is in The outcome ladder.

Get the same table for your machine

archie scan
archie stats ladder

archie stats ladder groups the index by model and repository and shows, for each group with enough sessions, how often it reached each rung. --min-n sets the floor (default 20), --json gives the rows. The same table is one MCP call away as stats_ladder, so your agent can read it too.

What this is not

It is not a model benchmark. The groups are one person's models on one person's repositories, and a rung says what evidence a session left, not whether the work was good. The opt-in aggregate export that would make cross-machine comparison possible is designed in archie-bench.md and not built.

apps/web/content/docs/learn/08-benchmarks.md on GitHub

Point it at your own machine.

npx -y agentworth@latest scan