--- title: 'In the News: November 13, 2026 (Extra 3)' description: 'Tessl says 75% of its agent-written PRs now merge without human review, and credits the standards its team writes for agents.' canonical_url: 'https://darkfactory.dev/news/2026-11-13-extra-3' markdown_url: 'https://darkfactory.dev/news/2026-11-13-extra-3.md' collection: news date_published: '2026-11-13T23:16:00-05:00' date_modified: '2026-11-13T23:16:00-05:00' --- # In the News: November 13, 2026 (Extra 3) Tessl reports that about three in four of its agent-written pull requests now merge with no human correction, and it publishes how it measured that. Also here: Anthropic's design guide for agents that run unattended across many sessions, and what a member of the Claude Code team says about the new coordinator in Claude Code Projects. ## 1. Tessl says 75% of its agent-written PRs merge without human review **[75% of Tessl's PRs now merge without human review](https://tessl.io/blog/75-of-tessls-prs-now-merge-without-human-review)** · Simon Rendon Arango and Colin Hallett, Tessl · Tessl blog, October 16, 2026 Tessl says its engineers open more than 1,200 agent-written pull requests a week in one monorepo, and that PRs that never merged plus review-fix rounds made up over half of its agent spend. It replaced its previous AI reviewer, which the post does not name, with its own product, Tessl Code Review, which checks each PR against the standards the team already writes for its agents. Since the switch, the post says, about 75% of PRs have merged "without a human correcting the agent or discussing the engineering." The comparison comes from one week in which both reviewers ran on every PR in the monorepo, more than 600 in total. Tessl counts 91% of its reviewer's findings as real defects, against 68% for the previous one, and 1.56 author-fixed bugs per PR against 0.75. A real defect had to be something that would misbehave in production; style and naming did not count. The post carries two caveats of its own. On an external benchmark with human-labeled defects, the two reviewers tied. And the previous reviewer was stronger on low-severity findings and caught some defects that Tessl's missed. It also reports that 8.5% of PRs went through more than five review rounds and used 47.4% of review loop spend. These are a vendor's figures about its own product, and nobody outside Tessl has reproduced them. **Why it matters:** The tie on the outside benchmark is the number to notice. On Tessl's own account, its reviewer pulled ahead only on the repo where it could check work against the team's written standards. For a team deciding whether agent PRs can merge unread, that points the effort at writing down its standards before it shops for a better reviewer. ## 2. Anthropic lays out the five parts of an agent that runs unattended **[Building always-on agents](https://claude.com/resources/articles/building-always-on-agents)** · William Hu, Anthropic · Claude blog, November 13, 2026 The guide says an always-on agent spends most of its time asleep: it wakes on an event, does one run, saves what it learned and stops. Around that loop it names five parts: an event inbox, timers, memory, permissions and an outbox. Event sources should write to a durable table rather than call the agent directly, because two events arriving together can start competing runs and a crash can lose an event. And "An inbox should only have one active worker at a time," so that two runs cannot duplicate work or overwrite each other. On memory, the guide says to keep only what the agent can't look up again, save pointers for the rest, and keep a progress file per task so the next run picks up where the last one stopped. For permissions it recommends a proxy outside the agent's environment that injects credentials, Claude Code's auto mode for higher-risk actions, and deferred tool calls that end the run and resume it when the user's answer arrives as a new inbox event. It gives one first-party cost figure: on a long-horizon coding benchmark, Opus 5.5 cost 78% less per solved task than Opus 5, against 40% less on typical workloads. The rest is design guidance, not a measured result. **Why it matters:** Agents on a schedule fail in the ways this guide plans for: two runs firing at once and overwriting shared state, or a crash that loses the event that started a run. A single worker per inbox, a durable event table and per-task progress files give a team running coding agents unattended a concrete list to check its own harness against. ## 3. A Claude Code team member describes the new Projects coordinator **[Hacker News discussion of the Claude Code Projects update](https://news.ycombinator.com/item?id=50492141)** · Boris Cherny, Anthropic (posting as bcherny) · Hacker News, November 13, 2026 Anthropic has updated Claude Code Projects, in which a coordinator hands work to threads that run in parallel. We could not find Anthropic's own announcement, so this item rests on comments Boris Cherny posted in a Hacker News thread about a third-party write-up, where he introduced himself as "Boris from the team." He wrote that the coordinator "delegates work to threads when needed," that threads "bubble up information and questions to the main thread," and that Claude runs threads in the cloud and checks out the relevant repos without being told which. On cost he wrote: "In AB tests and evals, we've found that the coordinator has <10% token overhead compared to running sessions directly." Anthropic has not published those tests. The thread had 62 points and 48 comments about three and a half hours after it was posted. One commenter said a single request spawned seven subagents and used up their plan's allowance in about ten minutes; Cherny replied that Claude is prompted not to use subagents unless asked, and asked for a bug report. **Why it matters:** Several commenters described coordinators they had built themselves out of skills and their own tools, and a vendor version gives those teams something to compare against. The only cost figure is Anthropic's own and unpublished, so a team should measure overhead and allowance burn on its own work before retiring a homegrown coordinator.