Loading today’s stories
Build1 distinct publisher3 min readUpdated
One rate limiter, a ten-million-token allowance and no termination criterion produced a diff 50 times the size of the job. The published fix only inspects the damage after the run ends.
The Engineer · Build desk
Compiled by The EngineerSomething wrong?How this is made
Start with what the published harness can and cannot see. It shells out to `git diff` after the agent's subprocess has already returned [15], so it reads the tree the agent left behind, not the path it took to get there. The behaviour that burned the afternoon leaves almost no trace in that tree, because edits that revert one another largely cancel out by the time the run ends [4]. The pendulum is the thing least likely to survive into the artefact the detector inspects.
The counter is also not counting what its name says. `edits_per_file` records a filename every time it meets a `+++ b/` line, then increments that file's tally for each following line beginning with a single `+` [16]. Those are added lines. The threshold applied to them is 8, and the author reads the same number as a file having been edited more than eight times [12][16]. Point it at the original task and the new limiter module clears eight added lines on its own [1], failing the run for oscillation it never performed [3].
The cost check has the same shape. Tokens are estimated at four characters each [13], measured over `proc.stdout` plus `proc.stderr` [14], which is what the agent printed to a terminal rather than what the model was billed for. An agent that reasons expensively and prints little passes the gate. One that narrates a straight-line fix fails it.
The diagnosis underneath is still the useful part: not the model and not the prompt, but a harness with no termination criterion beyond finish the task [10]. Turning that into a nonzero exit code [17] beats a silent bill. It is not yet a stop condition, because the spend has already happened by the time the check runs. Stopping has to sit between agent turns: hash each touched file after every step, and when a file returns to a state it has already held, end the run there.
One number is worth carrying out of this. The wrapper ships with a default budget of 150,000 tokens [12] against the ten million the author handed over before walking away [7]. The default is about 1.5 percent of the allowance, which means the allowance was roughly 67 times what the same engineer chose once he had to type the figure into a flag [2].
Provenance matters before anyone copies the file. The piece discloses that it was prepared as part of MonkeyCode's product outreach [6], the runner is MonkeyCode's own free-access tier [5], and the evidence is one task attempted twice [8]. Enough to make the mechanism credible. Not enough to tell you where your own thresholds belong.
Follow any of these and your For You feed starts watching them — no settings page required.
Ranked by verification strength, evidence, and original report placement.
Three hours after handing off the task, the author returned to a two-thousand-line diff.
The test suite was green, which the author says made the failure harder to explain.
The fix is a small Python harness of about forty lines that caps the run, measures the diff and fails loudly when the agent exceeds the cap or oscillates; the author says it is reproducible on any machine with Python and git.
estimate_tokens returns len(text) // 4, a heuristic of roughly four characters per token for code and prose.
Token usage is computed by running estimate_tokens over proc.stdout plus proc.stderr after the agent command completes.
The harness runs the agent command with subprocess.run, then after it returns runs git diff and flags files whose count exceeds max-edits-per-file.
Evidence-backed comparisons of source perspectives and observed adoption signals. Read the methodology
Which Builder, Operator, and Investor concerns the observed source mix emphasized—not a truth score.
Evidence, demonstrated adoption, hype gap, incentives, and confidence are assessed independently, each on its own current evidence. How these are measured.
Inspectable code, unverifiable incident
The only verifiable artefact is the inlined harness listing, whose defaults, estimator and diff parsing can be read directly. Everything load-bearing about the incident — three hours, a 2,000-line diff, fourteen edits to one file, a green suite, the reasoning trace, and the clean forty-one-line re-run — is a single author's uncorroborated recollection in a piece disclosed as vendor outreach, with no logs, repository or reproduction. Reading the published code also contradicts two of the article's own descriptions of what it does.
One self-reported user, no distribution
Adoption evidence is limited to the author's own two runs on MonkeyCode plus a script published as article text. No repository, package, release, version, download count, third-party operator or team deployment is cited, and the usage report sits inside a disclosed product-outreach piece.
Prescription outruns the published mechanism
The framing — assert budgets like test assertions, 'enforce them during the run', add an oscillation detector — is stronger than what the code delivers. All checks fire after the agent process exits, so a runaway run is fully paid for before the harness complains; token accounting is a len//4 character heuristic over visible output only; and the oscillation check counts added lines rather than edits, so the default threshold of 8 would misfire on the very forty-line change the article treats as the correct outcome. The headline diagnosis of budget-as-stop-condition is asserted from two runs of one task without controlled comparison. The article's own limitations paragraph partially offsets this by conceding the heuristic and the refactor breakage.
Disclosed vendor outreach
The article explicitly states it was prepared as part of MonkeyCode's product outreach, and MonkeyCode's free model access and free server option are introduced in the same opening paragraph as the incident. The narrative arc ends with the vendor's platform working cleanly once the author's own harness is added, so the promotional interest is direct; the disclosure itself is the mitigating factor.
Mechanism checkable, outcomes not
Confidence is moderate: the code-level assessment is firm because the full listing is supplied and its semantics are unambiguous, and the incentive read is firm because the disclosure is explicit. Confidence in the incident figures, the causal diagnosis and the reported improvement is low, since the cluster contains one publisher, one author, no artefacts and no independent corroboration.
build
Stop timing your GraphQL tests and start counting loader calls1 distinct publisher
build
Your 90% Cache Hit Ratio Is a Lagging Indicator. Alert on Cold Misses Per Key1 distinct publisher
build
The agent had permission. It did not have the current task.1 distinct publisher
build
An AI test suite hit 94% coverage and missed the one branch that mattered1 distinct publisher
Distinct publishers with included, body-backed reporting in this cluster.
dev.to
1 article · August 23, 2026