Proof-of-work for coding agents
Arxloop thinks through traces of the running code, where an agent working in a loop spends tokens. On the same model and the same tasks as Claude Code: up to 97% fewer tokens and up to 80% lower cost.
The same requests, word for word, to both, on the same model. Averaged over every task of a kind, bug fixes cost 65% less and builds 40% less. On a real open-source codebase, 7 of Arxloop’s 7 runs passed every check it was never shown.
| Agent | Runs | Cost a task | Hidden checks |
|---|---|---|---|
| WeaverClaude Sonnet | mean of 2 runs | $0.25 | 86/88 (97.7%) |
| Claude CodeClaude Sonnet | mean of 3 runs | $0.43 | 128/132 (97.0%) |
| Claude Codeasked to write tests | 1 run | $0.62 | 43/44 (97.7%) |
| What was asked | Arxloop | Claude Code | Arxloop’s cost | ||||||
|---|---|---|---|---|---|---|---|---|---|
| cost a run | tokens read | checks | runs | cost a run | tokens read | checks | runs | ||
| A real bug in a large codebaseA routing bug in Flask, a widely used open-source web framework, that its maintainers fixed after a release. | $0.49$0.39–$0.69 | 234k | 15/15 | 3 runs | $1.32$1.04–$1.51 | 3.7M | 15/15 | 3 runs | 63% less |
| A small bug fixA reported bug in a small project. | $0.03 | 10k | 8/8 | 2 runs | $0.17$0.14–$0.24 | 272k | 12/12 | 3 runs | 80% less |
| A real feature in a large codebaseA new setting in Flask, added on top of the fix in a new session. | $0.61$0.52–$0.77 | 256k | 36/36 | 4 runs | $1.07$0.75–$1.25 | 2.9M | 27/27 | 3 runs | 42% less |
| A database migrationMoving a database to a new schema, safe to run twice. | $0.24$0.21–$0.28 | 44k | 24/24 | 2 runs | $0.64$0.52–$0.70 | 1.3M | 36/36 | 3 runs | 62% less |
| A security auditFinding the security problems in a small service and fixing the serious ones. | $0.28$0.20–$0.36 | 75k | 27/28 | 2 runs | $0.66$0.53–$0.72 | 1.3M | 39/42 | 3 runs | 57% less |
| Missing testsWriting the tests for a module that has none. | $0.39$0.32–$0.45 | 90k | 5/6 | 2 runs | $0.51$0.46–$0.60 | 922k | 8/9 | 3 runs | 25% less |
| A slow scriptMaking a script ten times faster, with the measurements. | $0.31$0.30–$0.33 | 67k | 12/12 | 2 runs | $0.40$0.33–$0.50 | 751k | 18/18 | 3 runs | 22% less |
| A data clean-upCleaning a contacts file and reporting what was rejected. | $0.26$0.26–$0.27 | 45k | 10/10 | 2 runs | $0.24$0.19–$0.32 | 278k | 15/15 | 3 runs | 12% more |
Does a profound internal monologue actually help? Is it useful? Is it how humans used to fix code?
We used to gather a trace, form a hypothesis about what’s wrong, make the fix, get it wrong, learn from it, and try again — not narrate a profound monologue with the memory of a goldfish, staring at the same function for the fifth time going “ah, I see it now!” (again).
All that thinking costs tokens. Our harness skips the monologue — and still reasons.
A change is finished when it has been shown to work, not when the model says it is.
What it learned about your project on the last request is there for the next one.
What one person tells it once, the next person’s work respects without being told again.
Every run leaves a record you can read, question and undo.
Hand it recurring work on a schedule and read what it did in the morning.