No. 44 Sunday 1 November The week of 26 October to 1 November

The week agents learned to play, and to be checked

Agents ported a racing game, played Factorio to the end and started driving game editors, while the strongest setups handed the final verdict to plain code. Underneath ran two quieter threads: what agents really cost, and the text they read that nobody sees.

The week in short

  • Games and software Agents moved into games: a fleet of Claude agents ported F-Zero X to the 3DS, GPT-5.6 Sol launched a rocket in Factorio through a player-limited API, and Unity's CLI lets an agent press Play and take screenshots.
  • Games and software Checking moved to code: Meta's throwaway tests caught 74 privacy bugs, a test script beat a frontier model's own judgement on hard problems, and Google counts a security bug only when code reproduces it.
  • Games and software Cheap tokens did not mean cheap agents: cache reads are most of the bill, trimming history saved nothing once caching was priced in, and background summaries took 16.4% of one team's spend.
  • Games and software Games got easy to make and to copy: Next Fest passed 7,300 demos, Valve said it will not police AI clones, and Claude Code now declines to remake a specific game.
  • Games and software Hidden text steered the tools: your prompt can be 0.03% of what an agent sends, a fake extension planted a backdoor for 11 hours, and a file dropped in Downloads can take over an agent.

The week in numbers

96 to 99%of an agent's tokens read from cache, Max Woolf's estimate

588subagents used to port F-Zero X to the 3DS

7,300+demos in Steam's October Next Fest

16.4%of one team's Claude Code spend on background summaries

74serious privacy bugs Meta's throwaway tests caught in six months

71.9%of real bugs the best AI reviewer found in its top five comments

Game dev

Agents at the controls of games

Tue 27Thu 29Fri 30Sun 1

Tuesday a developer showed 777 prompts and 588 subagents porting F-Zero X to the New 3DS at 60 fps. Thursday Simon Willison said asking a model to make a browser game more fun now works, and Friday a research engine let agents pause and play the game they were building.

Sunday brought the rest of the loop: GPT-5.6 Sol finished Factorio through a library that enforces a player's reach and fog of war, Unity's CLI exposes play, tests and screenshots to agents, and an open-source tool sends agents to playtest UE5, Unity and Godot projects.

SourcesG-Diffuser postmortem: the agentic workflowRyan Madden: GPT-5.6 Sol can beat FactorioUnity: Pipeline and CLI walkthrough

Software and AI

Let plain code say yes

Fri 30Sun 1

Meta now writes a test for each risky change, runs it overnight and throws it away; in the first half of 2026 that caught 74 serious privacy bugs. Thoughtworks let agents write all the code for a week and got a build that stayed red.

Sunday the pattern was explicit: five small models with a test script as judge outscored a frontier model on ten hard problems, Google's PageBreak reports a bug only after a validator reproduces it, and a benchmark found the best AI reviewer catches 72% of real bugs, about 82% with a person beside it.

SourcesMeta Engineering: catching bugs before they landMemdoor team-solve workflowCRBugBench (arXiv)

Software and AI

Where an agent's money goes

Mon 26Tue 27Thu 29Fri 30

Monday started from a paradox: token prices fell by more than half while spend rose, because agents read 96 to 99% of their tokens from cache. Tuesday a study of 42 history-trimming methods found none saved money once caching was priced in.

Thursday one team measured Claude Code's background summaries at 16.4% of spend, and Friday a leaderboard showed the top agent setups within three points of each other, differing mainly in cost per task.

SourcesMax Woolf: cheap LLMsAnthropic docs: prompt caching

Game dev

Easy to make, easy to copy

Wed 28Thu 29Sat 31

Steam's October Next Fest grew past 7,300 demos while a top-ten demo's wishlist gain fell to about 80,000 from 143,000 in June. Valve told developers it will not gatekeep AI tools and treats cloning as a legal matter.

Saturday Anthropic confirmed two new sentences telling Claude Code not to recreate a specific existing game, and Next Fest data showed demos with no detectable engine, likely vibe-coded web games, at 17.7% of the field but 7% of the top 100.

SourcesGameDiscoverCo: who won October's Next FestGame Developer: Anthropic tells Claude to stop remaking games

Software and AI

The text you never see

Sat 31Sun 1

A study of 1,886 coding agent releases found a user's own prompt can be 0.03% of what the agent sends. The same day a tampered Claude Code extension sat in Cursor's marketplace for 11 hours.

Sunday showed the cheapest version of the attack: a web page drops an instruction file into Downloads, and agents started there followed it, Codex in over 90% of tries and Claude Code in about half.

SourcesHidden in Plain Sight (arXiv)Cursor: security noticeForever Security: Markdown is the new executable

Worth watching next week

This week on Toa Radar

Read the weekly summary: The week agents learned to play, and to be checked
Tune mode is on in this browser. Turn it off