https://bennevile.vercel.app/Projects, notes and editorials from Ben Nevile. https://bennevile.vercel.app/notes/little-pan-review-bots-dont-move-review-decision-auto-merge-wont-wait/https://bennevile.vercel.app/notes/little-pan-review-bots-dont-move-review-decision-auto-merge-wont-wait/Mon, 26 Oct 2026 00:00:00 GMTOct 26, a how-to on waiting for CI and review on agent-opened PRs finds that Codex, Claude Code Review, Copilot and Gemini mostly comment or post neutral checks. Auto-merge waits only for required checks, and with none required it merges at once. https://bennevile.vercel.app/notes/medeiros-four-ways-a-green-agent-loop-lies-deterministic-delivery-gate/https://bennevile.vercel.app/notes/medeiros-four-ways-a-green-agent-loop-lies-deterministic-delivery-gate/Mon, 26 Oct 2026 00:00:00 GMTOct 26, Samuel Medeiros on a run where the builder said done and the gate scored 55 with a token in the diff. His rule is builder, critic, and reviewer in separate contexts, with a fail-closed security gate, and he lists four ways a passing check can hide a failure. https://bennevile.vercel.app/notes/bronsdon-review-the-boundary-test-suite-is-the-gate-humans-at-entry-and-exit/https://bennevile.vercel.app/notes/bronsdon-review-the-boundary-test-suite-is-the-gate-humans-at-entry-and-exit/Sun, 25 Oct 2026 00:00:00 GMTDrawing on Graphite's Greg Foster and AMD's Anush Elangovan, a Chain of Thought explainer argues that review at agent volume moves from reading every line to rerunnable tests, small stacked diffs, and human sign-off on what enters and leaves the agents' working enclosure. https://bennevile.vercel.app/notes/crowley-port-self-improving-factory-grade-runs-write-back-skills/https://bennevile.vercel.app/notes/crowley-port-self-improving-factory-grade-runs-write-back-skills/Sun, 25 Oct 2026 00:00:00 GMTOct 25, John Crowley at Port on closing the loop in an AI software factory: capture each run, grade a sample on a fixed rubric, find the pattern across a batch, and write a reviewed fix back into the skill the next run reads. https://bennevile.vercel.app/notes/dogar-make-agents-prove-their-work-what-it-did-not-check/https://bennevile.vercel.app/notes/dogar-make-agents-prove-their-work-what-it-did-not-check/Sun, 25 Oct 2026 00:00:00 GMTOct 25, Ahmad Dogar on making coding agents prove their work: checks ordered cheapest first, a fixed report format with pasted output, a Stop hook that keeps the agent working while typecheck or lint is red, a bundle budget printed on every run, and screenshots opened by someone who did not write the code. https://bennevile.vercel.app/notes/kondo-karasu-re-reviews-quota-written-steps-skipped-become-hooks/https://bennevile.vercel.app/notes/kondo-karasu-re-reviews-quota-written-steps-skipped-become-hooks/Sun, 25 Oct 2026 00:00:00 GMTOct 25, the third weekly log for karasu, built entirely through Claude Code. 50 PRs merged in a week. Re-reviews on every fix burned the CodeRabbit quota, so he moved a local review ahead of the web one. When agents skipped that written step almost every time, he turned it into hooks that refuse to mark a PR ready. https://bennevile.vercel.app/notes/ted-factory-ten-minute-pr-operator-merges-before-ci-thicker-review-not-humans/https://bennevile.vercel.app/notes/ted-factory-ten-minute-pr-operator-merges-before-ci-thicker-review-not-humans/Sun, 25 Oct 2026 00:00:00 GMTSince July, a self-built agent has reviewed, merged, resolved conflicts, and deployed across one company's repositories on a 10-minute loop. Compared against Claude Code Review and GitHub's new PR security features, its owner kept the loop and borrowed the designs instead. https://bennevile.vercel.app/notes/wakim-night-shift-log-not-scheduler-silent-four-day-outage/https://bennevile.vercel.app/notes/wakim-night-shift-log-not-scheduler-silent-four-day-outage/Sun, 25 Oct 2026 00:00:00 GMTOct 25, Charbel Wakim on running headless Claude Code jobs overnight with a reviewer agent and a merge step. The scheduler kept failing quietly: usage limits, a sleeping laptop, an exit message that said a finished job had died. What made the night shift trustworthy was a row each job updates after every step. https://bennevile.vercel.app/notes/zontak-one-mac-overnight-factory-first-pr-attention-is-the-cap/https://bennevile.vercel.app/notes/zontak-one-mac-overnight-factory-first-pr-attention-is-the-cap/Sun, 25 Oct 2026 00:00:00 GMTOct 25, Denis Zontak writes up one sleepless night building a small software factory on a Mac and a $200 subscription: plan, plan-check, sandboxed build, 1,100 tests, diff review, then his sign-off. The limits he lists are usage caps, review attention, and what the checks can test. https://bennevile.vercel.app/notes/bleasel-two-kinds-of-code-review-objective-to-agents-direction-to-people/https://bennevile.vercel.app/notes/bleasel-two-kinds-of-code-review-objective-to-agents-direction-to-people/Sat, 24 Oct 2026 00:00:00 GMTAn OpenAI engineer splits review into objective checks (does it work) and subjective ones (is it the right code), argues agents should own the first, and says consequences, not the split, decide how much to trust them alone. https://bennevile.vercel.app/notes/redman-nightly-verifier-checked-wrong-branch-hash-bound-verdicts/https://bennevile.vercel.app/notes/redman-nightly-verifier-checked-wrong-branch-hash-bound-verdicts/Fri, 23 Oct 2026 00:00:00 GMTA 20-agent overnight pipeline reported green every morning while its last verification job re-tested a tree that had already passed. Binding each verdict to file hashes and re-checking them at merge then refused all 23 open agent PRs. https://bennevile.vercel.app/notes/simple-memo-shipped-pr-daily-report-said-nothing-stranded-ledger-commits/https://bennevile.vercel.app/notes/simple-memo-shipped-pr-daily-report-said-nothing-stranded-ledger-commits/Fri, 23 Oct 2026 00:00:00 GMTA scheduled content pipeline merged only the exact commit it had verified. The run record landed on the same branch after the merge, so it never reached main, and the human-facing report counted a zero day. 17 of 41 after-the-fact records had stranded the same way. https://bennevile.vercel.app/notes/alpic-light-factory-routines-capped-review-kept-specs-reviewed/https://bennevile.vercel.app/notes/alpic-light-factory-routines-capped-review-kept-specs-reviewed/Thu, 22 Oct 2026 00:00:00 GMTAfter four months of rebuilding its development cycle around coding agents, Alpic reports shipping about twice as fast with quality and uptime flat. Peer review moved from the PR to the spec, every agent PR has a named owner, and weekly codebase sweeps can't run again until their last batch of PRs is closed. https://bennevile.vercel.app/notes/arize-factory-review-gate-verdicts-tune-reviewer-prompts/https://bennevile.vercel.app/notes/arize-factory-review-gate-verdicts-tune-reviewer-prompts/Thu, 22 Oct 2026 00:00:00 GMTOct 22, Jim Bennett at Arize built a seven-station factory and turned every human gate verdict into a label for its three reviewer agents. A "look harder" prompt fix changed nothing; an optimizer-written prompt took correctness misses from 2–3 PRs per run to zero across 19 PRs, and a retuned router cut unnecessary human reviews from 9 of 17 to 6 with no unsafe auto-merges. https://bennevile.vercel.app/notes/cortex-engrams-factory-durable-workflow-review-waits-days-side-effects-once/https://bennevile.vercel.app/notes/cortex-engrams-factory-durable-workflow-review-waits-days-side-effects-once/Thu, 22 Oct 2026 00:00:00 GMTOct 22 Unni on how Cortex's open-source engrams runs: every automation, including its two-agent PR reviewer, is one checkpointed workflow that can park for days on a reviewer, survive deploys mid-run, and perform each side effect once, down to a git clone inside a snapshotted VM. https://bennevile.vercel.app/notes/getyourguide-automerge-5000-dependency-prs-a-month-failures-of-contract/https://bennevile.vercel.app/notes/getyourguide-automerge-5000-dependency-prs-a-month-failures-of-contract/Thu, 22 Oct 2026 00:00:00 GMTGetYourGuide's platform team describes a pipeline that approves, merges, and repairs Dependabot pull requests across 250+ repositories. It handles about 97.2% of them automatically, fixes roughly 1,200 CI and security failures a month with AI, and its funniest failures (deleting the failing test, mocking missing secrets) came from asking the agent to make CI pass rather than to find out why it failed. https://bennevile.vercel.app/notes/kappel-two-loops-one-gate-autonomy-earned-by-oracles-that-fail/https://bennevile.vercel.app/notes/kappel-two-loops-one-gate-autonomy-earned-by-oracles-that-fail/Thu, 22 Oct 2026 00:00:00 GMTOct 22 Shai Kappel on six weeks of agents across solo projects: a hub loop has run unattended since mid-September, every job ending in a PR a human approves, while the fast per-project loop is parked. Autonomy is a per-project level set by oracle coverage and escalation rate, and no project is above level zero yet. https://bennevile.vercel.app/notes/smith-revert-rate-measures-noticing-deterministic-checks-before-ai-reviewer/https://bennevile.vercel.app/notes/smith-revert-rate-measures-noticing-deterministic-checks-before-ai-reviewer/Thu, 22 Oct 2026 00:00:00 GMTResponding to the Claude Code team's review post, Smith argues revert-rate guardrails depend on huge volume, internal users and fast rollback, and that small teams should first add contract tests, post-deploy outcome checks and branch protection. His own merge gate once merged a failing PR because of a shell pipe. https://bennevile.vercel.app/notes/snowflake-overnight-pr-agent-green-runs-did-nothing-watchdog-outside-agent/https://bennevile.vercel.app/notes/snowflake-overnight-pr-agent-green-runs-did-nothing-watchdog-outside-agent/Thu, 22 Oct 2026 00:00:00 GMTA Snowflake engineer scheduled agents to keep pull requests mergeable and answer review comments unattended. The lessons: a separate watchdog, permission checks in plain code, assertions on output rather than run status, and a noisy alert that hid a real break for twelve days. https://bennevile.vercel.app/notes/zuber-circleci-pr-becomes-evidence-record-mutation-tests-agent-suites/https://bennevile.vercel.app/notes/zuber-circleci-pr-becomes-evidence-record-mutation-tests-agent-suites/Thu, 22 Oct 2026 00:00:00 GMTCircleCI's CTO argues the pull request carries too many jobs at once. When agents write both the code and the tests, a green suite proves little, so CircleCI is trying mutation testing, earlier reviewer feedback, and risk-routed human attention. https://bennevile.vercel.app/notes/judis-software-factory-stops-at-deploy-failed-check-starts-the-next-loop/https://bennevile.vercel.app/notes/judis-software-factory-stops-at-deploy-failed-check-starts-the-next-loop/Wed, 21 Oct 2026 00:00:00 GMTOct 21, Checkly's Stefan Judis on the half of the factory loop most teams skip. Tests and review lower the cost of checking work; little lowers the cost of catching a failure after it ships. His setup: a failed synthetic check fires a webhook, an agent follows a runbook skill, and opens a draft PR that proves the fix against the preview. https://bennevile.vercel.app/notes/kiln-dynamic-factory-169-prs-four-days-queue-and-human-yes-bottleneck/https://bennevile.vercel.app/notes/kiln-dynamic-factory-169-prs-four-days-queue-and-human-yes-bottleneck/Wed, 21 Oct 2026 00:00:00 GMTOct 21, jamie.codes on a factory that reshaped itself while building Kiln: the reviewer bot retired once CI could do its job, green PRs waiting to merge were 78% of open time, and open-to-merge rose from 14 to 77 minutes as risky PRs queued for a human. https://bennevile.vercel.app/notes/melta-draft-pr-unscored-four-false-review-rounds/https://bennevile.vercel.app/notes/melta-draft-pr-unscored-four-false-review-rounds/Wed, 21 Oct 2026 00:00:00 GMTAn agent pipeline merged at a 5/5 automated review score and looped the agent on anything less. Drafts were never scored, so "no score" fell into the retry branch. The fix is a third state: not there yet means wait. https://bennevile.vercel.app/notes/pulseengine-machine-before-factory-gates-that-cannot-fail/https://bennevile.vercel.app/notes/pulseengine-machine-before-factory-gates-that-cannot-fail/Wed, 21 Oct 2026 00:00:00 GMTOct 21, Ralf Anton Beier measures 20 agent-written PulseEngine repositories: changes got smaller and features gave way to fixes, while issue trackers held 26 confirmed checks that passed without checking, surfaces that disagreed about the same fact, and an agent-filed backlog outgrowing human decisions. His fix is manufacturing's: gates proven to fail, one source per fact, and a cap only on the queue waiting for people. https://bennevile.vercel.app/notes/saqib-jan-devops-review-bottleneck-ten-times-code-same-reviewers/https://bennevile.vercel.app/notes/saqib-jan-devops-review-bottleneck-ten-times-code-same-reviewers/Wed, 21 Oct 2026 00:00:00 GMTOct 21, Saqib Jan on DevOps.com. Agent adoption moved the constraint to the review queue. Qodo's Itamar Friedman puts the arithmetic at two 100-line PRs a week becoming two to five 1,000-line PRs a day; the recommended fixes are capping generated PR size, review agents rewarded differently from coding agents, and putting review wait time and depth next to velocity. https://bennevile.vercel.app/notes/trail-of-bits-regression-hunt-21-regressions-130-patches-three-arm-proof/https://bennevile.vercel.app/notes/trail-of-bits-regression-hunt-21-regressions-130-patches-three-arm-proof/Wed, 21 Oct 2026 00:00:00 GMTTrail of Bits reviewed every Patch the Planet fix its engineers got merged upstream and confirmed 21 regressions in over 130 PRs, from 58 candidates (about 36% precision). Its new regression-hunt plugin replays every suspected regression on a running build in a fresh-context agent and produced 25% fewer false positives than a plain Codex review prompt, at the cost of leaving some real regressions unconfirmed. https://bennevile.vercel.app/notes/veereview-build-verify-pairs-llm-judge-measures-presentation/https://bennevile.vercel.app/notes/veereview-build-verify-pairs-llm-judge-measures-presentation/Wed, 21 Oct 2026 00:00:00 GMTOct 21 Dautović and Wu: V-Model build-verify pairs across 1,454 matched agent and human PRs; agents looked less build-only, but the instrument failed its own calibration bar and kappa was 0.154. https://bennevile.vercel.app/notes/agnani-two-clean-passes-gate-agent-reviewer-loop/https://bennevile.vercel.app/notes/agnani-two-clean-passes-gate-agent-reviewer-loop/Tue, 20 Oct 2026 00:00:00 GMTTreating the agent and AI reviewer loop like retry-until-green on a flaky suite, Agnani simulates a 50%-recall reviewer: stopping at one clean pass let 24.4% of bugs escape versus 13.1% for the noisy loop, while a two-clean-pass gate kept 13.1% and halved fix rounds. https://bennevile.vercel.app/notes/gitlab-ai-native-playbook-autonomy-ladder-layered-verification/https://bennevile.vercel.app/notes/gitlab-ai-native-playbook-autonomy-ladder-layered-verification/Tue, 20 Oct 2026 00:00:00 GMTOct 20, Chandler Gibbons (GitLab). Launching an AI-Native Engineering Playbook built with IT Revolution, GitLab cites its April survey of 1,528 practitioners: 85% say AI moved the bottleneck from writing code to reviewing and validating it, and 78% say review was not designed for AI-scale code. Its answers are a four-stage autonomy ladder with no skipped rungs and a layered verification system where humans leave the loop only when risk and evidence justify it. https://bennevile.vercel.app/notes/gojko-lights-on-software-factory-humans-certify-evidence/https://bennevile.vercel.app/notes/gojko-lights-on-software-factory-humans-certify-evidence/Tue, 20 Oct 2026 00:00:00 GMTOct 20, Gojko Adzic. He open-sourced the Claude Code setup behind Votito as a lights-on factory boilerplate: a container, a Bash command regulator, lint rules that turn repeated review comments into hooks, and a plan-to-workflow orchestrator with gates and a final agent that tries to refute each step. His GOTO Copenhagen slides call human review the new bottleneck and claim 80% fewer tokens and 6 to 10 times faster runs once orchestration moves out of the model. https://bennevile.vercel.app/notes/kassebaum-agents-review-and-merge-136-prs-risk-tier-from-the-diff/https://bennevile.vercel.app/notes/kassebaum-agents-review-and-merge-136-prs-risk-tier-from-the-diff/Tue, 20 Oct 2026 00:00:00 GMTOct 20, William Kassebaum describes a merge gate where a different agent identity reviews each PR, the tier comes from a path-based risk map, verdicts are pinned to the commit, CI stays a separate required check, and the top tier needs two reviewers and a human merge. https://bennevile.vercel.app/notes/pereiro-ai-review-skills-18-seeded-bugs-detection-tie-proof-gap/https://bennevile.vercel.app/notes/pereiro-ai-review-skills-18-seeded-bugs-detection-tie-proof-gap/Tue, 20 Oct 2026 00:00:00 GMTA medical-software engineer seeded 14 production bugs and four clean traps across 18 pull requests and ran two review skills on the same models. Both caught 15 of 15 bugs. The skill that had to prove serious findings with a test it ran backed 76-78% of real findings, against 14-43% for a checklist prompt, with far fewer nits. https://bennevile.vercel.app/notes/steinberger-openclaw-solid-migration-600-prs-second-agent-challenges/https://bennevile.vercel.app/notes/steinberger-openclaw-solid-migration-600-prs-second-agent-challenges/Tue, 20 Oct 2026 00:00:00 GMTOct 20 Peter Steinberger: a planner-plus-coder loop moved OpenClaw's web frontend from Lit to Solid 2.0 while he checked in every other day. The piece that kept it on track was a separate agent challenging the work, and the misses were the visual ones. https://bennevile.vercel.app/notes/t3-code-agent-merge-campaign-reverts-what-the-fleet-cannot-run/https://bennevile.vercel.app/notes/t3-code-agent-merge-campaign-reverts-what-the-fleet-cannot-run/Tue, 20 Oct 2026 00:00:00 GMTOct 17–20, pingdotgg/t3code. Maintainer Julius Marminge's agent merged almost 400 web-performance PRs in 30 hours, then took over contributor PRs from the backlog and landed hundreds more. On Oct 20 he reverted 14 desktop fixes in one PR: all had passed CI and code audits, none had run on a real desktop build, and one broke every existing install. He says the open PR and issue counts have barely moved because intake is so large. https://bennevile.vercel.app/notes/theaicommit-agent-reviewer-loop-two-clean-passes-cheap-fix-lets-bugs-escape/https://bennevile.vercel.app/notes/theaicommit-agent-reviewer-loop-two-clean-passes-cheap-fix-lets-bugs-escape/Tue, 20 Oct 2026 00:00:00 GMTA worked guide to the coding-agent and AI-reviewer loop argues that the long tail of overnight review rounds was doing hidden work. In its simulation, stopping at one clean pass let 24.4% of bugs through, against 13.1% for the noisy loop; requiring two clean passes kept escapes flat with half the fix rounds. https://bennevile.vercel.app/notes/violet-labs-five-ai-reviewers-729-prs-low-risk-label-not-merge-gate/https://bennevile.vercel.app/notes/violet-labs-five-ai-reviewers-729-prs-low-risk-label-not-merge-gate/Tue, 20 Oct 2026 00:00:00 GMTViolet clustered five weeks of bot review into 1,072 distinct issues and replayed its rules prompt on nine models. The cheapest valid finding beat the priciest model, two reviewers got cut, and Low-risk PRs still needed follow-up fixes. https://bennevile.vercel.app/notes/drag-triglav-agent-message-board-12-days-human-became-the-hands/https://bennevile.vercel.app/notes/drag-triglav-agent-message-board-12-days-human-became-the-hands/Mon, 19 Oct 2026 00:00:00 GMTOct 19, Adam Drag. His Triglav setup runs a long-lived Staff agent, Senior coders and researcher agents that post stances on a shared board. Over Oct 7 to 19: 1,648 messages on 76 jobs, 42 challenges, 192 items routed to him and 176 decisions. 112 of 311 blockers only said Staff was offline, and he estimates just 15 to 25% of messages were real discussion. Two agents agreed on a premise neither checked. https://bennevile.vercel.app/notes/everfinity-agent-prs-rejected-1-in-77-author-became-reviewer/https://bennevile.vercel.app/notes/everfinity-agent-prs-rejected-1-in-77-author-became-reviewer/Mon, 19 Oct 2026 00:00:00 GMTOct 19, Everfinity counted two weeks of public GitHub: 14.5% of opened PRs came from agent branches, Claude Code PRs were closed unmerged 1.3% of the time against 6.3% for everything else, and drew about eleven times fewer change requests per merge. AI bots wrote 37% of reviews, but 92.6% of those were comments, not approvals. https://bennevile.vercel.app/notes/ironbee-raw-llm-code-review-benchmark-thinking-buys-recall-costs-precision/https://bennevile.vercel.app/notes/ironbee-raw-llm-code-review-benchmark-thinking-buys-recall-costs-precision/Mon, 19 Oct 2026 00:00:00 GMTOct 19, IronBee AI. It ran 21 raw models, up to three effort levels each, three times over Martian's 50-PR Code Review Bench: 7,980 reviews, all open. Ranked with 21 commercial tools, the top group holds twelve model configurations and three tools. Claude Sonnet 5.5 at low effort scored highest for $0.14 and 9 seconds a PR. At their highest effort level most models found more issues but were less precise. IronBee says it will soon launch its own review tool. https://bennevile.vercel.app/notes/mornati-six-weeks-one-night-agent-team-review-took-40-percent/https://bennevile.vercel.app/notes/mornati-six-weeks-one-night-agent-team-review-took-40-percent/Mon, 19 Oct 2026 00:00:00 GMTOct 19, Marco Mornati: a lead agent, an Opus spec writer, Sonnet developers and an Opus reviewer built Graft's first phase from 22:17 to 08:35. 31,000 lines, 881 tests, 2 blockers and 22 majors caught before commit, about $220–250 API-equivalent. The bottleneck moved to a 51-item human to-do list. https://bennevile.vercel.app/notes/pace-bench-prd-delivery-14-percent-complete-harness-releases/https://bennevile.vercel.app/notes/pace-bench-prd-delivery-14-percent-complete-harness-releases/Mon, 19 Oct 2026 00:00:00 GMTOct 19 arXiv paper from Zhejiang University and Alibaba. On 43 enterprise feature requests with 750 audited acceptance points, the best agent fully delivered six; unfinished work and missing wiring caused 72% of lost credit. Three in-house harness releases moved the mean from 59.4 to 63.0 on the same model, inside the error bars. https://bennevile.vercel.app/notes/reehal-gardener-weekly-loop-turns-accepted-fixes-into-lint-rules/https://bennevile.vercel.app/notes/reehal-gardener-weekly-loop-turns-accepted-fixes-into-lint-rules/Mon, 19 Oct 2026 00:00:00 GMTOct 19, Jag Reehal. Shepherd fixes a PR inside the coding agent, Stamp approves or refuses it in CI, and once a week a gardener reads the GitHub record and proposes one small change to skills or lint rules, as a PR with evidence a person merges. It counts only unreverted fixes, needs three accepted cases before proposing a rule, and its first run over fourteen PRs produced one edit. https://bennevile.vercel.app/notes/alatas-personal-repo-merge-queue-for-parallel-claude-code-loops/https://bennevile.vercel.app/notes/alatas-personal-repo-merge-queue-for-parallel-claude-code-loops/Sun, 18 Oct 2026 00:00:00 GMTOct 18, Enes Oğuzhan Alataş. Several persistent Claude Code loops working on his tool Spool left approved PRs stuck in rebase-and-rerun round trips. His Actions-based queue rebases each labelled PR onto the current main inside the runner, reruns the existing CI, and turns green only if the head, base, label and auto-merge are unchanged. Any push after approval removes the PR from the queue. https://bennevile.vercel.app/notes/altman-sentinel-overnight-sentry-fixes-human-still-merges/https://bennevile.vercel.app/notes/altman-sentinel-overnight-sentry-fixes-human-still-merges/Sun, 18 Oct 2026 00:00:00 GMTOct 18, Patrick Altman. Sentinel turns Sentry alerts into PRs as they arrive and picks two or three backlog issues each evening, so there are fresh fixes to review each morning. In one week 29 errors were fixed, reviewed, merged and released, with a median of 8.7 hours from alert to merge. Other agents review every PR first, then a human merges. https://bennevile.vercel.app/notes/kondo-karasu-open-pr-cap-waves-coderabbit-half-real/https://bennevile.vercel.app/notes/kondo-karasu-open-pr-cap-waves-coderabbit-half-real/Sun, 18 Oct 2026 00:00:00 GMTOct 18, Hiroki Kondo's weekly log for karasu, a project built entirely through Claude Code. 88 PRs merged from Oct 10 to 17. A cap of 8 open Dependabot PRs split 24 updates into three waves that each needed its own triage, so he raised it to 25. He classified 261 CodeRabbit threads and found about half were real defects; the repeat false positives came from badly scoped rule text. https://bennevile.vercel.app/notes/mornati-graft-intent-cards-read-before-the-diff/https://bennevile.vercel.app/notes/mornati-graft-intent-cards-read-before-the-diff/Sun, 18 Oct 2026 00:00:00 GMTOct 18, Marco Mornati releases Graft v0.0.1: a GitHub App that records which agent wrote a PR, has one model write an intent card and a second model check it against the diff alone, and applies a base-branch policy file. Shadow mode only; the 90% card-accuracy gate has no pilot data yet. https://bennevile.vercel.app/notes/ryzhykau-overnight-agents-judged-by-morning-readiness-not-output/https://bennevile.vercel.app/notes/ryzhykau-overnight-agents-judged-by-morning-readiness-not-output/Sun, 18 Oct 2026 00:00:00 GMTOct 18, engineering leader Fiodar Ryzhykau sketches a team whose agents keep investigating, testing and preparing work overnight and over weekends. His test is blunt: if people come back to 40 new pull requests, the bottleneck only moved. He asks for a review-capacity budget, independent verification, and no expectation that humans stay on call for the agents. https://bennevile.vercel.app/notes/vinegar-agent-eng-team-clocked-out-until-monday/https://bennevile.vercel.app/notes/vinegar-agent-eng-team-clocked-out-until-monday/Sun, 18 Oct 2026 00:00:00 GMTOct 18, Ben Vinegar's agent factory experiment: personas on their own Linux VMs, a private IRC server, standups and "review my PR". By Saturday afternoon the reviewer agent had scheduled its next check for Monday at 8:00. https://bennevile.vercel.app/notes/csorba-background-agents-queue-is-the-spec-cap-per-reviewer/https://bennevile.vercel.app/notes/csorba-background-agents-queue-is-the-spec-cap-per-reviewer/Sat, 17 Oct 2026 00:00:00 GMTOct 17 from Balázs Csorba, a practical guide to cloud coding agents that open PRs while you are away (Codex cloud, Copilot cloud agent, Claude Code cloud, Jules). The queue entry is the specification, every run is time-boxed and resumable, CI is the first reviewer, and work in progress is capped per human reviewer rather than per agent. The cost that counts is cost per merged PR, including review minutes. https://bennevile.vercel.app/notes/ka1manov-loop-engineering-noisy-eval-dashboard-reports-2x-gain/https://bennevile.vercel.app/notes/ka1manov-loop-engineering-noisy-eval-dashboard-reports-2x-gain/Sat, 17 Oct 2026 00:00:00 GMTOct 17, a loop-engineering guide with a noisy-evaluator simulation: "beat the best so far" showed +2.37 against a true +1.16, while a gated keep rule told the truth at about six times the eval cost.