Skip to content
self.md

■ SIGNALS // RADAR SIGNAL

review needs state

WorkBench shows workplace agents getting safer and more capable while Faros and GitHub show review queues absorbing the cost. The repair is inspectable state: evidence packets, isolated worktrees, and dashboard logs a human can audit.

Published Updated Read 4 min
■ [!] ON THIS PAGE ▼

self.md radar — 2026-06-15

the agent story got boring in the useful way: better task scores, more review debt, and state machines that make the work inspectable.

WorkBench says office agents are no longer just failing loudly. Faros and GitHub say the human review queue is taking the hit. TwinBI, wtdb, and data-review point at the repair: keep enough executable state around that a person can check the machine without rereading the whole universe.

1. WorkBench stopped looking like a toy lab

sources:

what happened: Olly Styles revisited WorkBench, the workplace-agent benchmark where GPT-4 completed 43% of tasks in March 2024 and took unintended harmful actions on 26% of them. The June 2026 run has Claude Opus 4.8 completing 89% of tasks with unintended harmful actions down to 2.5%.

The paper’s uncomfortable detail is that capability and safety move together on this benchmark: the models that finish more tasks also do less accidental damage. The remaining failures are not cute, though. Wrong-recipient email is still on the table.

why this matters: The adoption question shifts from “can the agent do office work?” to “which office actions are allowed to cross from suggestion into execution?” self.md needs that boundary as a product primitive, not a paragraph in a settings screen.

2. code review became the production choke point

sources:

what happened: Faros says median time in code review is up 441.5%, average review time is up 199.6%, and time to first review is up 156.6%. PRs merged with no human or agentic review are up 31.3%.

GitHub says Copilot code review has processed more than 60 million reviews, growing 10x in under a year, and more than one in five code reviews on the platform now involves an agent. The machine is already in the review loop. the awkward part is that humans still own the consequences.

why this matters: Generated code did not remove the expensive part; it moved the bill to trust. A personal AI OS that ships work needs evidence packets, risk tiers, and a named human owner for the load-bearing merge.

3. review needs executable state

sources:

what happened: TwinBI ties an LLM agent to an executable BI dashboard state and a unified interaction log. With the same backbone agent, exact-match accuracy moves from 43.3% to 63.3%, partial-credit accuracy from 48.3% to 70.8%, and timeout rate from 40.0% to 10.0%.

The smaller tools rhyme with it. wtdb gives each git worktree its own copied database and env wiring. data-review reruns affected pipeline lanes, diffs the numbers against a blessed baseline, and checks whether the movement matches declared intent.

why this matters: The useful pattern is not “more memory.” it is state you can replay, diff, isolate, and blame. once agents touch dashboards, databases, and business numbers, chat history is not enough proof.

  • Agentic Code Review — the best long read today on why review, not generation, is the scarce part of agentic coding.
  • Early-Stage Prediction of Review Effort — a useful older paper for triaging agent-authored PRs before a senior engineer burns an hour.
  • agentic-os — a small governance repo that treats plan, build, review, test, and ship as evidence gates.
  • Orchestra-o1 — omnimodal orchestration research that fits the same pressure: the agent has to decompose work before it can be trusted with it.

left on the table

agents code-review workflow personal-ai