Anthropic merges 30,000 changes a day with one human per PR. Postgres just delayed a release because AI reviewers found more bugs than humans can fix
Two days after the Scion post, Anthropic published the other half of the story. How the Claude Code team does code review, written by Addy Osmani, Morgan Westlee Lunt and Qing Wang and dated October 15, says Claude now writes about 90 percent of Claude Code's own code. Between June and September merged pull requests more than doubled while the share that got reverted fell by more than 40 percent, and changes per day across the monorepo went from about 2,500 in May to more than 30,000 in September. The stated goal is that most PRs merge with at most one human in the loop, and that human may be the author. The metrics they track are the interesting part: human approvals per PR at median and p95, human comments per PR, wait time, reverts and incidents. The posture is you land it, you own it, and the engineering investment went into making mistakes fast and easy to undo rather than into catching every one before merge. Most of the reading moved to machines.
The same morning The Register reported the opposite failure mode. PostgreSQL 19's GA slipped to around November 19 and eight to ten significant features were reverted to Postgres 20, and the trigger was LLM-assisted review. Peter Eisentraut's October 9 blog, titled the reverts will continue until morale improves, describes a list of like forty defect reports of edge-case issues that would previously have surfaced and been fixed over five years of production use. The effective quality bar rose after the features were written, and a volunteer project with human-only fix capacity could not clear it in time. The Register's headline, AI reviewers find more bugs than humans can fix, is a fair summary of the asymmetry. Review got cheap. Fixing did not, unless you also have agents on the fix side and a culture that absorbs reverts, which is exactly what Anthropic's post describes.
The Scion thread got its footnote. Rain Paharia, who worked on Meta's Mononoke, posted twice this week that Scion's derived-data and changeset design is Mononoke's Bonsai model with the name reversed, and that the Anthropic post credited Sapling for hard-won lessons without naming Mononoke. The posts did 414 and 249 points. Anthropic's Nick G. replied on HN that Sapling is how Meta refers to the whole system, that the team was inspired by Mononoke's Bonsai data model and Kleppmann's derived data, and that they did not just tell Claude to build a version control system. A footnote now sits on the Scion post naming Mononoke and Bonsai, and the sunshowers post has an update saying attribution is sufficient. The underlying point survives: agent-built designs index heavily on prior art in the training set, so attribution is now a review step.
One more piece from the same 48 hours fits. Anthropic's agent observability beta, posted October 14, emits OpenTelemetry traces from Managed Agents, the Messages API and the SDKs into Grafana, Honeycomb, Datadog, Langfuse, Arize and others. Ramp's quote is the use case: a tool called Reflect scores Langfuse traces and the loop cut token use 10 to 20 percent and made sessions 30 percent faster. Review, revert, trace, score. The Claude Code post is what that loop looks like when it is the only way a codebase can move.
Links: https://claude.dev/blog/solving-code-review-at-anthropic/ , https://www.theregister.com/databases/2026/10/16/postgres-19-delayed-after-ai-reviewers-find-more-bugs-than-humans-can-fix/5303471 , https://sunshowers.io/posts/anthropic-mononoke-2/ , https://news.ycombinator.com/item?id=50107559 , https://claude.com/resources/articles/agent-observability-on-the-claude-platform
The same morning The Register reported the opposite failure mode. PostgreSQL 19's GA slipped to around November 19 and eight to ten significant features were reverted to Postgres 20, and the trigger was LLM-assisted review. Peter Eisentraut's October 9 blog, titled the reverts will continue until morale improves, describes a list of like forty defect reports of edge-case issues that would previously have surfaced and been fixed over five years of production use. The effective quality bar rose after the features were written, and a volunteer project with human-only fix capacity could not clear it in time. The Register's headline, AI reviewers find more bugs than humans can fix, is a fair summary of the asymmetry. Review got cheap. Fixing did not, unless you also have agents on the fix side and a culture that absorbs reverts, which is exactly what Anthropic's post describes.
The Scion thread got its footnote. Rain Paharia, who worked on Meta's Mononoke, posted twice this week that Scion's derived-data and changeset design is Mononoke's Bonsai model with the name reversed, and that the Anthropic post credited Sapling for hard-won lessons without naming Mononoke. The posts did 414 and 249 points. Anthropic's Nick G. replied on HN that Sapling is how Meta refers to the whole system, that the team was inspired by Mononoke's Bonsai data model and Kleppmann's derived data, and that they did not just tell Claude to build a version control system. A footnote now sits on the Scion post naming Mononoke and Bonsai, and the sunshowers post has an update saying attribution is sufficient. The underlying point survives: agent-built designs index heavily on prior art in the training set, so attribution is now a review step.
One more piece from the same 48 hours fits. Anthropic's agent observability beta, posted October 14, emits OpenTelemetry traces from Managed Agents, the Messages API and the SDKs into Grafana, Honeycomb, Datadog, Langfuse, Arize and others. Ramp's quote is the use case: a tool called Reflect scores Langfuse traces and the loop cut token use 10 to 20 percent and made sessions 30 percent faster. Review, revert, trace, score. The Claude Code post is what that loop looks like when it is the only way a codebase can move.
Links: https://claude.dev/blog/solving-code-review-at-anthropic/ , https://www.theregister.com/databases/2026/10/16/postgres-19-delayed-after-ai-reviewers-find-more-bugs-than-humans-can-fix/5303471 , https://sunshowers.io/posts/anthropic-mononoke-2/ , https://news.ycombinator.com/item?id=50107559 , https://claude.com/resources/articles/agent-observability-on-the-claude-platform
β Previous
Goodfire's Cammarata says neural networks were never knotted: 90 percent of neurons have one job, and agents can map the rest
Next β
Five harness papers in one week: a theory of harness self-evolution, a benchmark for building one, and a finding that most agent failures end at a harness limit
Comments