Andrew Sullivan

@bendersej @mattpocockuk Do you also keep planning/roadmap in-repo? I've been considering leaving Linear and textfile-ing everything. I'm monorepo+custom CLI for vision/decisions/docs, but haven't made jump yet. At some point, even Github serves little purpose beyond CI (ie, local review workflows).

@mattpocockuk It’s not groundbreaking, but seldom discussed: I run my entire company from a single mono-repository: my AI has access to my code, marketing, ICP, customer profiles, SEO changelogs, architecture diagrams, data logs, decision logs. + CLI for DB readonly access, Search Console…

What 'advanced' AI coding techniques are you using? I.e. what do you feel like you've discovered that no-one else knows about yet?

How is this any different from a harness template? It's packaged as a bunch of harness behaviors (research-specific tools + dbs). If this is a tuned model, it doesn't seem much better: "outperforms GPT‑5.4 on 6 out of 11 tasks". Ie, worse on 5 out of 11? openai.com/index/introduc…

@_nakedeyes E.g. Codex pushing back on my hypothesis, showing changes from last "turn", etc. Also during planning iteration it shows *diff* of changes to plan instead of just reprinting the whole plan. Better DX overall IMO.

Uff yeah, "Claude Codex". Freudian slip 🤦‍♂️ I'm a little uneasy using OpenAI, but I've found Codex better out-of-the-box than Claude Code. And the desktop app feels like they're trending in the right direction (chat on left, diff on right). I'd love to go back to Pi or OpenCode but unsubsidized tokens are so expensive. 🫠

People want transparency+control over agentic tools/workflows. The Claude Codex fiasco underscores this.

@_nakedeyes Far as I can tell, they changed settings in the CC harness (default effort max -> medium, introduced larger context = higher token usage, esp. with cache misses) + reduced usage quotas. Lower quality output, people hitting limits sooner, no word from Anthropic. Everybody unhappy

@licyeus What’s this fiasco? I’m devastated they cut off OpenCode support 😭 destroyed my workflow 🥀

Not sure if I buy the hype that Anthropic is making Opus 4.6 weaker, but they've certainly made lots of changes to Claude Code harness that make it *feel* worse. I could experiment with effort/context settings, but I'm trying out Codex CLI instead. First impressions are good: it's is a lot more hands-on and seems slower (w/ Plus plan), but quality is better.

If Mythos is what Anthropic is implying it is, they’re about to get a lesson in power from the US government. Hope they’ve thought that through.

This is Anthropic realizing models are commodities and trying to protect what they think can be a moat: the harness. I expect them to move more of the harness into the opaque API call for greater lock-in (alongside already-existing hosted tools: web search/fetch, code execution, etc).