Product Radar Edition 03 · 30 Oct 2026
Edition02

T3 Code and Haiku move in, Codex goes to Hold

My daily work moved to Opus 5.5, Haiku 5.5 and T3 Code, and I dropped Codex. I also stopped my own review of agent work and now rely on automatic checks.

Anthropic's Claude models Opus, Sonnet and Haiku go from largest to smallest. Fable is Anthropic's model for hard, long tasks. Opus 5.5 now does almost all my work. Haiku 5.5 moves to Adopt. It does the small jobs, such as a quick check of each change my agents make. Fable 5.1 and Sonnet 5.5 move to Trial. I use them only for a few narrow jobs. Codex is OpenAI's coding agent, and GPT-6.1 Sol is its model. Both go to Hold. I cancelled my Codex plan on 5 October and last used it on 13 October. The models fell short for my work.

T3 Code moves to Adopt. It is the app where I start and watch my agents. This radar calls such an app a harness. T3 Code now also starts my scheduled agent runs. MLX Whisper (speech-to-text on my Mac), Nano Banana 2.1 (images) and TypeSafe Jev 1.13 (a cheap yes-or-no checker for text) move to Adopt as my defaults. Computer use, where an agent clicks and types in Mac apps for me, comes back in Trial. One new Infrastructure entry, Local models for agent work, goes to Assess.

I rewrote the practices for people who build products. They go from 12 to 8. The main change: I stopped my own review of agent work. Automatic checks now decide what ships, as Guardrails, not manual review explains. Mask personal data before a model sees it moves to Adopt. Four practices are new. Under Not on the radar, I list 9 tools that I left off, each with its reason.

  1. Check every path through a user flowNew in TrialNew this week: found 16 bugs in four flows.
  2. Give agents narrow keysNew in TrialNew: the lesson from a host's security notice.
  3. Guardrails, not manual reviewNew in TrialNew: replaces four narrower entries.
  4. Local models for agent workNew in AssessTested this week: no gain on a 16 GB Mac yet.
  5. Scheduled agent runsNew in TrialNew: how my scheduled agent runs work since October.
  6. Claude Haiku 5.5Trial → AdoptDefault for quick checks and for clicking through Mac apps.
  7. Computer use on macOSHold → TrialBack in use through my agents' own screen tool.
  8. MLX WhisperTrial → AdoptTranscribes every voice memo on my Mac.
  9. Mask personal data before a model sees itTrial → AdoptCode enforces it whenever an agent reads support mail.
  10. Nano Banana 2.1Trial → AdoptMy default image model.
  11. T3 CodeTrial → AdoptNow starts my scheduled agent runs.
  12. TypeSafe Jev 1.13Assess → AdoptChecks quotes on my site twice a day. Now testing it as a cheap first reviewer.
  13. Claude Fable 5.1Adopt → TrialFew sessions now: Opus 5.5 took over the daily work.
  14. Claude Sonnet 5.5Adopt → TrialTwo narrow jobs now: long searches and release review.
  15. CodexTrial → HoldPlan cancelled in October. The models fell short for my work.
  16. GPT-6.1 SolTrial → HoldStopped: I cancelled my Codex plan in October.
  17. AGENTS.md filesRetired from AdoptFolded into Company brain to keep the practices short.
  18. Compiler feedback for agentsRetired from TrialFolded into Guardrails, not manual review.
  19. Cross-model reviewRetired from TrialFolded into Guardrails, not manual review.
  20. Guard hooks for agentsRetired from AdoptFolded into Guardrails, not manual review.
  21. Model tier per taskRetired from AdoptFolded into the Claude model entries.
  22. Models select, code verifiesRetired from TrialFolded into the Gemini 3.8 Flash entry.
  23. Parallel agents in worktreesRetired from AdoptStill my daily practice, but no longer news.
  24. Shipping untested agent codeRetired from HoldRestated as a positive rule in Guardrails, not manual review.