Claude Fable is a genius, but expensive af. GPT-5.5 is a flawless reliable operator, affordable with a max plan. Combine them and you get the best of both worlds. Use Fable as the architect, Codex as the builder, the repo as the memory and yourself as the judge. You spend minimal tokens on Fable while getting maximum value in the form of a plan/audit. Codex simply executes on the plan which might take hours, which is fine on a max plan.

quick tip If you're using multiple harnesses in the same repo: - Make AGENTS.md the shared master file for all your harnesses claude/codex/pi/grok/etc. - Give claude a one line bridge in the CLAUDE.md file to use AGENTS.md instead. If you don't claude won't read agents.md, but claude.md and you'll get inferior output for your setup. - Next session: run '/context all' and confirm it shows under memory files.

Google AI sub is the best value per $ subscription around. IF you make it so your agents can use it. I got tired of paying API prices for Google models my sub already includes. Agents now hit the consumer apps instead. Image, video, Deep Research, NotebookLM.

The no fluff version of what 'Jev' is and enables: Jev is a new type of model. It's extremely fast and cheap. Small decisions can cost 40 to 400 times less and run 20 to 200 times faster. Jev is not replacing GPT, Grok or Claude. Frontier models still reasons, write and plan. Jev can't: - produce natural language - code - reason But he could do the small calls models do for every task. The simple version of what that means: An LLM is the thing that can reason and generate. It reads context and writes an answer. But an agent also makes hundreds of tiny decisions like: - is this message urgent - does this action match the plan - should I escalate - is this output safe to pass on. Right now almost every one of those goes through the same slow, expensive model. So we batch them. We check less often than we should, because every check costs a real turn which would make it slower and more expensive. Jev is a very different type of model. It cannot generate text. Think of it like a calculator instead of a writer. I expect there to be tons of consequences for agents. To make it more tangible, two examples of what I'd imagine this changes for agents: 1. Improved autonomy. A long-running workflow could score “is this state still inside the framwork, or do we need a human now?” on a tight loop and only page someone when confidence drops, instead of either proceeding and hoping things go well or stopping for a full model turn every time it's unsure. 2. While an agent is mid-refactor, it could fire a cheap “does this change still match the accepted plan?” check on every meaningful edit instead of waiting for one slow LLM review at the end. What excites me most is that it is so insanely cheap and fast for it's intelligence level, that it changes what an agent can afford to check. Which opens up all kinds of possibilities that weren't feasible before and I can't fathom yet. Fingers crossed it's not just marketing hype.

Stop reviewing your agents' work with a model from the same provider! Or worse, asking the agent to check its own work. Here's what to do instead: I spoke to a lot of AI operators over the last weeks and surprisingly many did either of the above. Which is terrible for output quality. Think about how humans review impactful work: - review the work yourself - ask a colleague to review - review colleagues findings and fix Same with AI, the same model provider reviewing the main agents' work has the same blind spots, so it won't find all the issues. Instead what you should be doing once the main agent completes its work -> spawning a new session powered by a model provider different from the agent that did the work and asking it to review the work of the main agent. This results in far superior output quality. My 2 review flows For any impactful agent work; - Main agent (codex/grok) works on task. - Hook fires when the main agent finishes. - Hook initiates 1 or 2 agents in new sessions powered by different model providers (fable/deepseek). They run a 'review' skill over the main agent's work. - review agents' output is delivered to main agent which confirms their findings and fixes the relevant findings. - Fixes are controlled by the same 1/2 review sessions, if deemed fixed appropriately the main agent continues with the next step or delivers its final output to me. (See the attached diagram for the actual flow) For any minor work that needs checking, but you don't want to waste time/tokens I simply scale down the reviewer models: - Spawn cheap and quick review agents on high reasoning levels Fable -> Sonnet, Deepseek-pro -> deepseek-flash. Examples A few from my own logs. September 1, after a rename pass across the repo, I told the main session to "spawn tens of cheap gpt-5.6-luna subagents at max reasoning to double check your work. Once you have findings, confirm them for yourself." Result: 22 auditors, 46 raw findings, 12 unique, 7 real, all fixed. The big one (10 auditors found it independently) was five broken links to a renamed file. The main session's own "all clean" check had missed it, because its link checker skipped dot-directories. The main model checked itself and passed itself. The day after that, during a website bakeoff where five models each built a batch of sites, a Claude operator reported three video deliveries as "delivered and verified". A Grok investigator I had running on the side found all three were one local placeholder file copied three times, same md5 checksum (a file fingerprint, identical only if the bytes are identical). Verified first-hand. The "delivered and verified" status was wrong, and the model that wrote it would never have caught it. The thing to remember is, a reviewing agent from the same model class has the same weak spots. You cannot check a Claude model with another Claude model. If Claude is your main, you check it with a Codex model and a Grok model. In my experience it works far better. A few more from the logs, so you know it's a pattern and not one lucky catch: - September 3, same bakeoff: every seat reported its site's effective frame rate. I had the orchestrator measure each one in a live browser. sol claimed 60.18, measured 60.3. opus claimed 75.0, measured 75.0. Only then did I believe the reports. - September 5 and 6: five sub-agents from five families (deepseek-v4-pro, claude-fable-5, claude-opus-5, gpt-5.6-sol, grok 4.6, all on medium reasoning) reviewed what the main model had built. When the reviews came in: "three things I had shipped were wrong." Grok found a bug in the main model's own verification command. The main model's own words: "I generalized a single data point." - September 8: four cheap annotators labelled the same text and came back with 22%, 85%, 2% and 51% for the same question. Four different answers to one question means the measurement is broken. One Grok read of the full corpus found the issue none of them saw. Later that evening Grok reviewed the plan and the Claude session wrote: "Four of my positions were wrong, and one of the refusals was cowardice." fml Demand "proof" "Give me proof this works" is a prompt I use a lot. The problem is what comes back. If the same agent wrote the fix, ran the test and rendered the screenshot, all I have is the agent's claim with a picture attached. Would I accept it? Yes, I would. That's exactly how you get fooled. So the proof has to be something the agent couldn't have written itself. - production logs it can only read, not edit - a /version endpoint that returns the deployed commit, plus CI, plus the actual deploy - a checksum of the file it says it delivered - a frame rate measured in a real browser - for anything about look and feel, I click around myself Ensure the agent automatically delivers you the proof. Don't put it in AGENTS.md. Make it part of the reviewing skill that you invoke or that fires on a hook, so every agent checks the same way instead of inventing its own definition of "done". Three more things to watch out for: 1. an agent that writes both the fix and the test will make them agree. 2. A checker that judges by a proxy (log size, exit code, "tests passed") misses the failure. 3. And the most common mistake: a reviewer from your main model's family will nod along, because it has the same blind spots. The rule In an ideal world you use one model as your main, and one or two different model classes for checking, depending on how important the task is. Right now my mains are Sol/Astra from Codex, and Grok 4.6. The reviewers are DeepSeek (Flash/pro) and Claude Fable. And they need your main model to confirm their findings before implementing anything, because the reviewers have blind spots too and most of their findings are noise or very minor nitpicks (46 raw, 7 real). I build it into the pipeline as standard practice. Costs It's annoying. One or two extra models per check, more tokens, and sometimes the reviewers are wrong in confident ways (a Sonnet judge in one of my sessions preferred the version with the vaguer number, twice). My reading is still the bottleneck. I have written about that before. Always remember, the thing AI is best at is persuasion. You can see that in its output. It can extremely confidently say that it has done something correctly, but it could be completely fabricated or taken a shortcut to achieve the desired outcome. Hence the review agent/judge is so critical.

I absolutely love Grok 4.6 and @orca_build. But the "ping", "ping", "ping" on every task completion (not just final output) of Grok is driving me mental. I fixed the issue, but once Orca updates it overwrites my fix and I have to fix it again. Please fix permanently, Orca chads.🙏

Wow, sounds like Astra is a new breed of model. Capable of actually thinking from first principles vs being a great memoriser... ARC-AGI-3 is a test made of little games. No words, no numbers, no hints. You just get dropped in and have to figure out the rules yourself, like a video game with no tutorial. A few months ago, every frontier AI model tried this and failed almost completely. Less than 1 out of 100 tries worked. A human can solve basically all of them, first try, no practice. OpenAI's new model just solved 999 out of 1000. Most models are good because they've seen a million examples before and just remember the pattern. This test was built so memorizing doesn't help at all. There's nothing to memorize. So doing well means it actually figured things out on the spot, like a person would. So what does that actually unlock... Right now, AI is only really good at stuff it's already seen a ton of. Ask it to write an email, it's seen a billion emails. Ask it to code in Python, it's seen way more Python than any human ever will. It's basically a genius at things that already have a big pile of examples to learn from. But the second you hand it something genuinely new, with no examples anywhere, it usually falls apart. A weird piece of internal company software nobody's ever documented. A one-off lab machine with buttons that don't match anything online. A robot arm with a sensor setup somebody just built last week. There's no training data for any of that, because it didn't exist yet. An AI that can score 99.9% on a test designed to have zero shortcuts is an AI that might actually handle those situations. Not because it memorized the answer, but because it can poke around, notice what happens, build a little model in its head of how the thing works, and figure out the goal without anyone explaining it. Sounds like something that is actually able to invent new shit, something that all previous models have struggled with.

In the not too distant future you'll sit down on the couch. Open your TV and the 3 custom-made personalised movies your AI picked out for you are presented. You don't even have to think about what you feel like watching anymore as the AI will be able to predict this before you think about it. Just absolutely wild... higgsfield.ai/@higgsfield.st…

Cursor, terminals, VS Code… I tried a lot of apps for agentic work. Working in multiple repos on VPSs simultaneously meant 2–3 windows and constant flipping. Super annoying. @orca_build is the first thing that fixed that for me. What I needed was one workspace that treats multi-repo, multi-session agent work as the default job. One day on it already beat everything else I tried before.

Sometimes I still try to fix a weak agent by giving it more tools. Usually with the opposite effect and the only thing that goes up is the bill and the noise. What I check first now is the boring harness stuff: - early context rewrites - always-on tools that never fire - uncapped dumps Stable prefix and fewer always-on tools beat another skill pile more often than I want to admit.

I've been looking for the perfect interface to do agentic work. I finally found it and it's bliss. Over the past year I've used Cursor, CMUX, Ghostty, vscode, claude code and codex apps. Generally I disliked anything that tried to pull codex/claude out of the terminal to 'improve the UX'. I was looking for an interface where I: - Could run 4-8 parallel sessions - Have an explorer view of my files and can view any file type - Am able to read rendered markdown files - Natively browse - Use any frontier model I stuck with vscode, as it had everything I wanted when running agents in the terminal. Even though it wasn't really built for it and it's a memory hog. It wasn't great though since; - I couldn't work in multiple repo's without opening multiple vscode instances (which further ate into memory usage). So often I had 2-3 instances open and switched between them while working I meant to try out @orca_build for a while and finally gave it a try. Wow, just sheer perfection! Everything I was looking for and more. Truly built from the ground up for agentic work. I can finally work in multiple repos in the same interface, no more flipping through multiple vscode instances! It even has: - Built-in design tools like: 'grab page element', 'annotate page element' and 'draw on screenshot' function for any kind of visual files. I used plugins and external apps for this before, but Orca has it natively built in which is awesome. Even the annotations on this screenshot were drawn with it. - CC/Codex usage tracking There's even more I haven't tried out yet, but even with just the 'bare bones' it's already superior to everything else I've tried. gg @orca_build and thank you for the great app🙏