DocIQ

Tried this recipe today, and flow-next was a natural test bed for it: its specs are already built to be handed to any model (acceptance criteria, edge cases, a reviewed implementation plan), and cross-model review loops and orchestration are wired in for both interactive sessions and autonomous runs. Swapping models per stage took zero glue code. One task, hidden oracle-graded test suite, planted-bug review eval, three repetitions per review arm at matched reasoning efforts. Plan: scored by two blind judges from different model families, identical ranking. Fable 5 came out at 38/40 average against Sol high's 31, and pushing Sol to xhigh only moved it to 32.5. All plans were sound and caught every trap. The gap was concreteness and scope discipline: Fable wrote actual test cases with expected values and stayed exactly in scope, Sol left a checklist and added unrequested behavior. Build: as many of us have been saying for what feels like years now, the spec + implementation plan carries it. Given a frontier-authored plan, all 11 workers, from Luna medium to Terra xhigh, followed it perfectly: 39/39 on the hidden suite, every single one, and every one kept the plan's brute-force oracle test. The separation showed up elsewhere. Wall-clock depends heavily on tool-loop iterations, and the more capable tiers needed fewer of them. The test suites they wrote alongside the code split by family too: the GPT tiers turned 27-61 of the spec's edge cases into assertions, Claude, Grok and Cursor's Composer 78-155, regardless of effort setting. Effort bought reasoning time and nothing else. Review is where models separate. Three runs per model at matched effort, scored against four known bugs (three planted, one latent): Sol high caught all four in every run, at the fastest mean time. Luna xhigh matched that recall at 2.5x the time. Grok 4.5 caught the three classic bugs each time, and missed the subtle latent one each time. One false positive across 27 review runs, zero hallucinated findings otherwise. And the shape of the review follows the prompt, not the model: ask for a bare defect list and that is what strong models return; ask for the minimal fix, blast radius and severity per finding, and Sol and Luna both deliver a complete, correct fix plan. Fable volunteered that shape unprompted -- which, held to the standard I just applied to the plans, is also scope creep. The useful kind, this time, but the lesson stands: spell out the artifact you want back. That is where actionability comes from. So: build is a commodity when the spec is strong. Plan and review are not. Put your intelligence there, especially if you want a more autonomous style of agentic engineering. flow-next.dev

Spent a couple of days convinced my Codex app was broken. It was 4,633 chat threads in an 887 MB sqlite db that the UI re-renders all the time. Pruned them. 5 MB, back to instant. ~/.codex/state_5.sqlite for the record.

Off to the 🏖️ in Italy for a week tomorrow, remoting to codex set up of course. Got some OSS work done before leaving, flow-next.dev now also works in @cursor_ai and @grok build (in addition to the claude code and cursor)

Getting some good traction with this

For my Mac friends here that don't want to use cloud solutions for Dictation and Meeting Transcriptions: I was too annoyed with other options, so I built dettivo.com. One-time purchase, on-device, agent first, fine-tuned and fully optimised (59ms voice to insert).

Looking for an AI-pilled keynote speaker. Germany (in English), late October. Paid. ~130 CTOs and CEOs etc. across a PE portfolio in the room. Brief: where this is going at the operating-margin level. Inspiring, not job-loss. Tag someone. Any labs interested? @OpenAIDevs @AnthropicAI

I'm shit at marketing.. Some things I've been working on aside from my upcoming orchestration thingy: gno.sh: added a publishing option to gno (local retrieval stack, perfect for for your @karpathy style vaults) dettivo.com: the best local mac dictation and meeting transcription app out there (am biased) Might get around to doing some actual marketing eventually. For now I just use them myself daily and I like them

This is the way, running a very similar setup that i've been meaning to post about for a couple of days now... I'll get around to it eventually but: I also started out with my Obsidian Vault for implementing @karpathy's llm-wiki idea, it is basically my command center for ALL day-to-day knowledge work, ie. - ingesting anything interesting via keep.md - creates all my work documents / decks that are published as artificats to our SharePoint (boo) via onedrive sync -- the markdown files remain the source of truth - can read all 8 of my mail accounts via gogcli and mail.app (no graph API access for our exchange accounts) - research via /last30days I never use Obsidian anymore, only the vault and their file format (backlinks, frontmatter etc), I replaced it with gno.sh which doubles as both the semantic retrieval layer and file viewer/editor. You can also use qmd for the retrieval part if you prefer.

Finally sat down and rebuilt Flow-Next's Codex support from scratch. Native plugin, pre-built agents with sandbox modes and nicknames, skill metadata for the Codex UI, the works. Was tired of the "clone repo and run a shell script" install. Now it's just /plugins → install, same as Claude Code. 20 agents, the whole workflow. One hook short of full Ralph parity on Codex (they don't intercept Edit/Write yet) but everything else works. Cross-model reviews can still go through @repoprompt or Codex CLI depending on what you like (hint: use RepoPrompt, it's awesome). github.com/gmickel/flow-next

LFG👀

Agentic Engineering: The Boring Stuff #1 Quick reminder for everyone that didn't catch my post last night. I want to start a series of Boring Tips / War Stories for people trying to scale agentic coding to teams/entire companies, etc. This is the first post, so I'll set the scene: part of my day job is helping our portfolio companies roll out 'AI-native Software Development Lifecycle', which is just a fancy way of saying that I help them roll out coding agents, new methodologies and processes across the company so they can speed up their cycle times. To be clear, these are not startups that are already on the cutting edge, these are mid-market EU-based companies in the 50-500M revenue range that have maybe, but only maybe heard of Copilot before. Nearly all of them practice some form of Agile theatre and are, quite frankly, not set up in a way to just install Codex and start churning out code, they simply can't leverage it. Just introducing the tooling would possibly result in like a 10% uplift, but, as everyone here knows, so much more is possible. When I work with these companies, I provide them with a curated guide that I update almost daily and that has grown to around 100k words. I then spend a lot of my time working through the guide, methodologies, processes with 4-8 people from the company who can then implement it and scale it to the rest of the company. Now that's out of the way, today's topic: So, a few weeks ago I said Agile is dead. Here's what I meant. The #1 reason teams don't see speedups after introducing agentic coding or whatever you want to call it, is of course not the model or the tooling. They always make the mistake of not changing how they work, ie. they just plug in Codex or Claude Code and still perform whatever bastardised form of Agile theatre around it. But: - Sprints assume human-speed iteration. Agents iterate in hours. - Story points assume implementation is the effort. It isn't. - Standups assume you even need daily syncs. - Refinement assumes human cognitive limits. What actually works: - Holistic specs, not fragmented tickets. How much faster is your team/company going to be shipping if you're feeding 2 tickets to a developer every two weeks? - Measure cycle time and not made-up velocity. - Scale up your planning and review capacity (more on this in future posts). That's where the real bottlenecks are now. - The Spec defines done, not some magic sprint boundary. McKinsey: teams that restructure hit 6x the delivery gains but 70% of companies haven't changed a single role definition despite adopting AI tools. Put another way, the gains are there to be had but almost nobody is doing the thing that unlocks them, they are bolting AI onto unchanged orgs and wondering why nothing changes. They are running a 2026 engine on a 2001 operating system. What did your team actually change? I'll cover the best responses in this series.

Creating @DocIQ_io's new site today (we have two incredible new products launching very soon). When I say creating, I'm talking about a specialised Ralph Loop that is just iterating on the design, copy, remotion components all day until Opus, Gemini and Codex are happy with it.

Flow-Swarm + @OpenAI Codex, running autonomously for 25 hours straight. My own take on @cursor_ai's FastRender experiment. Obviously I can't run 2000 agents for a week. Had to scope it way down. Bought two additional fresh ChatGPT Pro accounts just for this. Peak concurrency: 10-11 agents (beginner numbers but still). Orchestrating GPT-5.3 Codex (high) (@OpenAIDevs cooked) has been impressive. Stays on task for hours. Doesn't drift. It's just a phenomenal model. Not going to say what it is yet. But there's an entire category of software that's been coasting for years because building a good version is (was!) genuinely hard. So everyone ships something half-broken and charges enterprise prices for it. It's about 5% away from acceptance gates I defined right now. The incumbents should be nervous.

Flow-Next with Opus 4.6 automatically works on tasks in parallel. I didn't change anything.

The upcoming @DocIQ_io uses AI Elements heavily, highly recommend! That and @convex of course