No. 42 Sunday 18 October The week of 12 October to 18 October

The week proof became the bottleneck

Agents now write most of the code at the companies that build them. This week's stories were about what that leaves for people: checking, choosing the risks, and keeping the keys safe.

The week in short

  • Games and software Anthropic, Meta and Uber all described the same shift: agents write and test, and human time moves to review, risk and judgment.
  • Games and software A cheap model plus a test harness matched an expensive model alone in one paired experiment; checking, not writing, decided the result.
  • Games and software Agents rebuilt a shipped game from its binary, and a decompilation index shut down: shipped code is now readable.
  • Games and software Two very large open-weight models arrived in one week, from Meta and DeepSeek.
  • Games and software Engines are changing their foundations: Unity 7 moves to CoreCLR, RimWorld left Unity, and Godot 4.8 reached beta.

The week in numbers

30,000+daily code changes across Anthropic's monorepo in September, up from about 2,500 in May

99%of a shipped shooter's functions rebuilt by agents from its binary

3 to 10accepted runs out of 10 for a cheaper model once a test harness was added

8 weeksfor Meta's Forum team to reach feature parity with Facebook Groups

40.7 ptsof task state lost between an agent's run and its reports, in one study

Software and AI

Verification is the new bottleneck

Sat 17Sun 18

The Claude Code team now sends every pull request through parallel AI reviewers and sets human review by risk tier. Meta's Forum team reached feature parity in eight weeks with agents running the build and test loop and engineers approving evidence.

The costs showed too: PostgreSQL 19 slipped after AI review found more defects than its developers could fix, and a paired experiment showed a harness rescued a cheap model while only slowing an expensive one.

SourcesClaude Code blog: solving code reviewEngineering at Meta: How Meta built Forummartinfowler.com: Are Harnesses Worth the Cost?Peter Eisentraut

Game dev

Epic goes AI-first, and builds tools that way

Wed 14Thu 15Fri 16

Epic moved teams to AI-first production and said the gains would reach everyone on Unreal. Its new launcher was vibe coded, and it is rebuilding Easy Anti-Cheat for Linux with Fortnite as the pilot.

SourcesGame DeveloperWccftech

Games and software

Shipped code is readable now

Thu 15Sat 17

A team of agents decompiled a shipped shooter, rebuilding 99% of its functions with 83% byte-exact, kept honest by a verifier they could not edit. Later in the week decomp.dev closed its public index, citing AI-finished projects.

Sourcesmomo5502Simon Willison

Game dev

Engines swap their foundations

Wed 14Fri 16Sun 18

Unity 7's alpha runs on CoreCLR instead of Mono, though Unity says faster iteration is a longer-term goal. RimWorld left Unity for its own engine and reports 2 to 4 times better speed. Godot 4.8 reached beta with a long rendering list.

SourcesUnity: CoreCLR scripting updateGodot 4.8 beta 1

Software and AI

Keys, sandboxes and the supply chain

Thu 15Sat 17Sun 18

npm's Trusted Publishing was shown to trust a package name rather than its repo, Chainguard put coding agents in throwaway machines with signed results, and OpenCode said provider keys saved in its console may have leaked.

SourcesOfficeChai: OpenCode breach

Worth watching next week

This week on Toa Radar

Read the weekly summary: The week proof became the bottleneck
Tune mode is on in this browser. Turn it off