---
title: 'In the News: September 24, 2026 (Extra 2)'
description: 'An open source tool reviews agent-written code through diagrams and decision logs instead of raw diffs.'
canonical_url: 'https://darkfactory.dev/news/2026-09-24-extra-2'
markdown_url: 'https://darkfactory.dev/news/2026-09-24-extra-2.md'
collection: news
date_published: '2026-09-24T16:13:00-04:00'
date_modified: '2026-09-24T16:13:00-04:00'
---

# In the News: September 24, 2026 (Extra 2)


A four-person team shipped an open source IDE built to make agent-authored pull requests reviewable, with a decision log and a diagram layer standing in for the raw diff. A public benchmark shows how much of that review burden comes down to harness choice rather than model choice. And one infrastructure engineer makes the case for scheduling a fleet of agents the way distributed systems already schedule unreliable machines.

## 1. An open source IDE for reviewing what an agent decided, not just what it changed

**[Show HN: Whiteboard, an open source IDE for thoughtful software design](https://github.com/devdotfast/whiteboard)** · Sid, Alex, Ketan and Milan, devdotfast (Y Combinator W26) · Hacker News, Sep 24, 2026

Whiteboard is an open source desktop IDE, built on the VS Code codebase, that gives a coding agent an SDK to draw sequence diagrams, entity relationship diagrams and quoted trace excerpts onto a canvas linked to the underlying code. Clicking a diagram element jumps straight to the code it describes. A semantic, AST-aware diff viewer, written in Rust, summarizes large generated functions as pseudocode and collapses routine changes such as tests and documentation by default. A separate decision log lets an agent record why it made an autonomous choice, queryable later against its own trace. The founders say teams at Salesforce and Modal already use it as a review gate for architecture and spec-level changes, a claim that comes from the founders themselves and has no outside confirmation yet. The tool is MIT licensed; a hosted version with multiplayer review is planned. At 114 points and 35 comments about three hours after posting, one of the founders wrote: "As more PRs were merged without our understanding, we felt a cognitive debt begin to seep in, until it became difficult for us to even contribute to the system."

**Why it matters:** Most teams answer the volume problem by reading less of the diff. Whiteboard answers it by moving the review up a level, to the diagram and the stated decision, so a reviewer can check what an agent meant to do before checking every line of what it did.

---

## Also this cycle

- **[Browser Agent Evals](https://www.stagehand.dev/evals)** · Browserbase (Stagehand) · updated Sep 3, 2026. A public leaderboard runs 15 model-and-harness combinations against one fixed benchmark. Holding the model constant, Claude Opus 5 alone ranges from 71 to 78 percent accuracy depending on which harness runs it, claude code, deep agents, eve or fx, which is evidence that harness choice moves accuracy on its own. The cheapest combination, GPT-6 Luna on Codex, reaches 74 percent accuracy at $0.0225 a task, against $0.51 a task for the top-accuracy combination, Claude Opus 5.5 on Claude Code at 89 percent. Browserbase invites others to run the same harness against their own models.
- **[A Million Agents Is a Distributed Systems Problem](https://www.instacloud.com/blogs/a-million-agents-is-a-distributed-systems-problem)** · Tony Chang, InstaCloud (InsForge) · Sep 2, 2026, recirculated on Hacker News Sep 24. Chang argues a fleet of agents should be scheduled and recovered like a distributed system, not treated as a bigger brain. He cites a 180-configuration Google Research study in which a centralized orchestrator held error amplification to 4.4 times versus 17.2 times for agents working independently, and ACL 2026's Silo-Bench, where agent teams working on distributed reasoning tasks fell to zero percent success at 50 agents on the hardest tier and stayed there at 100. His prescription: keep goals, plans and checkpoints durable outside any single agent, so a coordinator can fail and be replaced without losing the work.