Загрузка видео...
Не удалось загрузить видео
Introducing flow-1, our new model trained with RL to find errors in agent traces. It matches GPT-6-sol in trace intelligence while being 23x cheaper. It also costs 25% less to run than GPT-6-luna. flow-1 finally makes it possible to monitor and understand every agent run, without sampling. 1/6
244,417 просмотров • 8 дней назад •via X (Twitter)
Комментарии: 20
flow-1 runs inside Signals agent - a harness deeply optimized for agent trace analysis. It operates like a coding agent, treating each trace like repo and each span like a file. As an input it takes a user prompt, which could as general as "find logical errors in this trace", an output schema, and a trace. It greps span content, reads tool results and returns structured findings.
On our Signals benchmark of 523 difficult agent traces across coding, legal, customer support, healthcare and more, flow-1 matches GPT-6-sol in analysis quality. Sol flags more traces. flow-1 is more precise about which ones contain real failures.
We compared costs by running the models inside the Signals agent on traces of various difficulty that have under 100k total tokens. flow-1 delivers GPT-6-sol-level analysis while processing 23x more traces per $1!
We started with supervised fine-tuning on synthetic investigations of realistic workflows across software, legal, customer support, CRM and more. We then trained with RL inside the Signals agent to improve tool use, schema adherence and investigation of complex traces. Here’s how the training data breaks down by industry.
With flow-1, production traces become a source of better evals, better agents and insight into what users need. Read the full post for results, training details and more examples of what flow-1 finds:
Amazing launch, Robert and team!! Love what you're building and the direction Laminar is heading!
Hell yeah, congrats Robert! This looks really cool.
How to get AGI: 1. Train a model that can generate response 2. Train a model that can detect errors at the previous model. 3. Repeat 1~2
Very interesting! won't it lose generalization ability if you do this?
It’s not a general model, it’s purpose built for trace analysis and works primarily in our Signals harness
excellent idea if the boundary is well defined!
congrats guys!
analyzing 100K agent traces of 50k tokens each costs under $250 with flow-1. crazy
Congrats on the launch!
super cool!
yummy @ClevrPwn
Is error detection a classification problem?
Why would I use this over OAI or Anthropic? Lmao
Beautiful video btw
Incredible tracing solution
![Official OpenAI tip for Codex: stop burning GPT-6.1 Sol context on tasks GPT-6 Luna can swarm, while GPT-6 Astra sits idle put Astra on call as a dedicated adversary codex --model gpt-6.1-sol --agent adversary GPT-6.1 Sol acts as the lead architect, scoping architecture and validating pull requests GPT-6 Luna teammates take isolated git worktrees and hammer out implementations at medium effort GPT-6 Astra monitors the peer-to-peer bus as a dedicated adversary, only stepping in at three critical moments: → before a contract locks: does the frontend payload match the backend schema? → when a test breaks twice: are we patching the bug or just hiding the symptom? → before calling done: what attack surface or edge case did everyone overlook? GPT-6 Luna builds. GPT-6 Astra challenges. GPT-6.1 Sol merges and ships Jev engineering runs the identical pattern one layer down: mechanical decisions that need no reasoning (which file to open, which tool to invoke, retry or abort) execute in 16ms, so the frontier models only wake up when execution paths actually diverge Plan on high. Delegate on medium. Keep the adversary on peer-to-peer call. - the full team setup > GPT-6.1 Sol on high coordinates the session and synthesizes PRs > teammate-ux builds client state inside a separate git worktree > teammate-back writes endpoints and handles Redis atomic locks > teammate-adv runs invariant fuzzing and catches replay vulnerabilities > all teammates communicate directly over peer-to-peer messaging without coordinator drag > tasks.json manages file-level mutex locks so nobody steps on another worktree paste the setup and this prompt into Codex CLI below: "Configure my Codex workspace for autonomous agent teams: 1. Audit ~/.codex/agents and .codex/agents for teammate profiles fitting ux, backend, and adversary > Generate definitions only for missing roles > Pin model: gpt-6-luna, model_reasoning_effort: medium for workers, and model: gpt-6-astra for adversary > Assign each worker an isolated git worktree branch to eliminate file conflicts 2. Enable agent teams in ~/.codex/config.toml: > Set [agents] max_concurrent_threads_per_session = 8 > Lock main session model to gpt-6.1-sol with model_reasoning_effort = 'high' > Set approvals_reviewer = 'auto_review' 3. Initialize tasks.json in repo root with atomic file locking, allowing teammates to claim subtasks and resolve contracts directly over inter-agent messaging without coordinator intervention 4. Add a team rule to AGENTS.md: teammates must request an adversary security review before opening a pull request for the lead session Show every configuration diff first. Do not apply edits until confirmed" ↳](https://image.24vids.com/tw-2106127135278612573/amplify_video_thumb/2106126443818160128/img/um3FcaWLOtHQKOfu.jpg)
![[Launch Week 1/5] Introducing Agent Identities Automation breaks at signup, login, or 2FA because agents rely on external accounts instead of their own. With Agent Identities, each agent gets a real email, phone and 2FA flow so it can sign up, log in and keep running without human intervention.](https://image.24vids.com/tw-2000605864445763702/amplify_video_thumb/2000603713950396422/img/I-6gAu74peXH15-8.jpg)