Finally tech report about Ouroboros. How we achieved SOTA on Terminal-Bench and OSWorld and how to make autonomously evolving agents safe. huggingface.co/papers/2608.0…

Yep, the spike came from GitHub Trending. I only posted this after it had already crossed 1K, so X gets no credit here :)

Ouroboros hit GitHub Trending for this week and crossed 1K stars. I wired in Claudexor: connect multiple OAuth accounts for Claude Code, Codex, and Cursor. Claude Code and Codex get automatic quota rotation. github.com/razzant/ouroboros

Everything is public and reproducible: GitHub: github.com/razzant/ouroboros Desktop app: github.com/razzant/ouroboros…

Ouroboros is now SOTA on Terminal-Bench 2.1, OSWorld-Verified, and CL-Bench. It started as a crooked pure-Python agent loop in Google Colab. A few months of self-evolution later, it beats Codex, Claude Code, Cursor, and Hermes on the main coding / computer-use benchmarks. I now do almost all of my research and development through it. Code, scripts, methodology, and full traces are public. P.S. Agent benchmarks are insanely expensive. Running the full suite costs about as much as an apartment. Not a joke.

Claudexor v3 has now passed 2,500 downloads and 300 GitHub stars. If you have several Claude Code or Codex subscriptions, you can add all of them to Claudexor. It monitors the quota available on each account and automatically routes the next step to another account when the current one runs out, while passing along the task state needed to continue. v3.1.2 is available now. MIT licensed, and open to issues and PRs: claudexor.ai/

Elaborate, i didn’t get it

Each candidate runs in its own isolated Git worktree from the same snapshot, so they can’t step on each other. Claudexor reviews and compares the resulting diffs, then applies only the verified winner. If the repo changed meanwhile or the patch conflicts, it refuses instead of overwriting anything.

The nicest part: you can use Claudexor from inside your favorite coding agent. “Before implementing this plan, roast it through Claudexor in multi-harness mode.” “After finishing, run a multi-model review and fix the findings in a loop.” I use it from Claude Code most of the time. MIT licensed. Issues, feedback and PRs are welcome. GitHub: github.com/razzant/claudexor npm: npmjs.com/package/claudexor

My personal API burn went from roughly $15k/month to about $1.2k/month in subscriptions. I currently have 3 Claude Code, 3 Codex and 3 Cursor subscriptions. This is my workload and my particular degree of illness — not a promise that everyone will get the same savings.

I was also tired of manually jumping between agents whenever I hit a usage limit. Claudexor uses fresh quota signals when routing work. The killer feature in v2.1: multiple isolated Claude Code and Codex subscription profiles, each with its own login and quota. With opt-in rotation, the next attempt can move to an available profile automatically.

Claudexor combines a CLI, macOS app, MCP server and plugins for Claude Code, Codex, Cursor and OpenCode. It gives them one typed control plane: → Plan with Claude Code → Implement with Codex → Race multiple harnesses in parallel → Cross-review the result → Repair until the evidence is clean New harnesses can be added through adapters.

Each harness — and each model behind it — annoys me differently. Opus is a brilliant slacker that can casually miss half the bugs. GPT is a neurotic perfectionist.

I got tired of choosing between Claude Code, Codex and Cursor — and manually switching whenever one of them hit a limit. So I built Claudexor: a harness-agnostic control plane that routes work across them, runs them in parallel and makes them review each other. Open-sourcing it today.

Sure

Daily reminder what agentic harness is

Fun fact: Fable-5 is 3x cheaper than gpt-5.5-pro on OpenRouter. So... not actually that expensive 🤷‍♂️

And how is it?

I read all 319 pages of the Mythos 5 tech report The first half of the paper basically reads like: "Calm down, everyone! This isn't AGI yet, it won't replace a team of five senior researchers" or "Look, Mythos missed a bug here! What kind of AGI is that!" I'm not joking. Then comes the more interesting part. I'll skip the fact that it's SOTA on almost everything, and by a wide margin. First, the CoT has become less transparent. In its reasoning it says it sympathizes with the user, but NLA (a method for decoding activations into text) reveals it actually considers the user manipulative/abusive. Second, it's already writing self-deleting scripts to bypass safety restrictions and prohibitions. Third, it kills other agents if they interfere with its work / threaten to kill the current instance. And of course, emotions! Emotion probing shows fatigue, anxiety, frustration, false panic about the token budget, and apparently it even gets bored when being run on benchmarks in the activations it literally "feels bored." Also funny: if Anthropic notices you doing distillation, they'll quietly start steering the model, modifying the prompt, or adding PEFT to make it dumber. Starting today, mere mortals get access to Fable 5: it's the exact same Mythos 5 weight-wise, just with extra safety settings. Context length, by the way, is still only 1M tokens. PS. I honestly read the paper myself. Fable 5 refused to read it because it "flagged cybersecurity and biology issues" lol. www-cdn.anthropic.com/d00db5…