No. 46 Sunday 15 November The week of 9 November to 15 November

The week the checks moved outside the agent

Again and again, the work that held up was judged by something the agent could not edit: a table of tests, real AWS, a proof checker, a fuzzer's score. Meanwhile Unreal opened its editor to agents and small, cheap games kept finding big crowds.

The week in short

  • Games and software Checks moved outside the agent: tests kept as tables, an emulator proven against real AWS, a sidecar that confirms failures, and Google's falsifiers that agents cannot edit.
  • Games and software What an agent reads mattered: a stale test command was followed in 78 of 80 runs, and old TODO comments added 70% more steps.
  • Games and software Cheaper models took over more steps: Gemini 4 Hydrogen edits while Carbon plans, and a small classifier replaced a frontier model at about a 44th of the cost.
  • Games and software Unreal Engine 5.9's preview lets Claude Code, Codex, Cursor or OpenCode drive the editor over MCP, and brings back a browser target.
  • Games and software Small games won big: a solo $8 co-op game passed an estimated 250,000 copies, and a fuzzer won NetHack in minutes.

The week in numbers

78 of 80runs that followed a stale test command in an instruction file

$8 / $40Claude Fable 5.2 price per million tokens in and out

70%more steps when agents chased old TODO comments, per Google

442CVEs credited to Anthropic in seven months, per VulnCheck

79%of 480 NetHack games won by Antithesis's fuzzer

250,000+copies of Crack, a solo $8 co-op game, GameDiscoverCo estimate

Software and AI

Keep the judge outside the agent

Thu 12Sat 14Sun 15

Thursday a Thoughtworks engineer argued for tests kept as data tables an agent cannot quietly change, and AWS described building a local emulator with agents only after each test had passed against real AWS. Saturday Momentic rated just 9.6% of 1,000 agent-written unit tests useful, and Google released a sidecar that confirms each production agent failure before reporting it.

Sunday Google Research described /teamwork, where many agents explore and only results that pass a falsifier survive: a Lean proof checker, unmodified tests or a simulator. A Rust port of its AV1 decoder matched the C++ bit for bit and ended 20 to 30% faster, by Google's account.

SourcesMartin Fowler: tests as dataStack Overflow: your coding agent should not grade its own homeworkGoogle Research: /teamwork

Software and AI

What the agent reads decides what it does

Mon 9Wed 11Fri 13Sun 15

Monday a study of 732 runs found agents follow a stale line in an instruction file almost every time, 78 of 80 runs for a moved test command, and rarely say so. Wednesday Google showed old TODO comments send agents on unasked detours, with 70% more steps.

Friday a paper found most broken prompt caching fails silently, and Sunday a developer measured 52 connected MCP tools at 24,600 tokens of descriptions per turn, against 640 for four focused tools, with fewer right first picks.

SourcesStale instructions study (arXiv)Google Developers: agent distractionDEV: MCP servers draining your context

Software and AI

The right model for each step

Tue 10Wed 11Fri 13Sun 15

Tuesday Claude Fable 5.2 arrived at $8 in and $40 out per million tokens, with effort level mattering as much as the model name. Wednesday Amazon swapped a frontier model for a small classifier on a labelling job, about 44 times cheaper by its own figures.

Friday Google let one Antigravity task plan with Gemini 4 Carbon and hand edits to the cheaper Hydrogen, for 24% less quota. Sunday Perplexity released Decider v1.3, an open model for routing and judging in one pass.

SourcesAnthropic: Claude Fable and Mythos 5.2Antigravity: Gemini 4 HydrogenPerplexity: Decider v1.3

Game dev

Engines open the door to agents

Mon 9Wed 11Fri 13Sat 14

Monday Frontier3DBench found agents can build props in Blender and Unreal but score about 30 of 100 on stateful gameplay. Wednesday Epic's Unreal Engine 5.9 preview, mostly written by agents in about a week, added an editor panel that runs Claude Code, Codex, Cursor or OpenCode over Unreal MCP, plus an experimental browser target.

Friday a five-person studio that left Fortnite showed its Defold pipeline: a 4-second boot, 6,500 tests and an MCP server for agents. Saturday Antithesis's fuzzer won 79% of 480 NetHack games by saving, branching and scoring game states.

SourcesUnreal Engine 5.9 MegaJam previewDefold: Snowfall Games and PortentumAntithesis: NetHack

Game dev

Small games, found by players

Sat 14Sun 15

Saturday Talking Tom's maker reported 357% more visits referred by AI assistants after six months of work, and the EU's Digital Fairness Act became a formal proposal aimed at subscription traps and hidden fees.

Sunday GameDiscoverCo profiled Crack, an $8 co-op game one developer made in nine months, at an estimated 250,000 copies, and shared a playbook for reaching Chinese players on Steam. A MIGS talk showed AI-disclosed games make up 40% of Steam releases but about 13% of reviews.

SourcesGameDiscoverCo: CrackPocketGamer.biz: Talking Tom and AI referralsPocketGamer.biz: MIGS and PGC Montreal

Worth watching next week

This week on Toa Radar

Read the weekly summary: The week the checks moved outside the agent
Tune mode is on in this browser. Turn it off