Laminar flow-1: The RL-Trained Model That Finds Hidden Errors in AI Agent Traces
Laminar flow-1 is a new model from Laminar (YC S24) that reads AI agent traces and flags the failures nobody noticed. Released on October 2, 2026, it matches GPT-6-sol in detection quality on Laminar’s own benchmark while costing 23x less to run. That makes it affordable to check every production run instead of a small sample.

Key Takeaways
- Release: Laminar announced flow-1 on its blog on October 2, 2026, and presented it on Hacker News on October 19, 2026.
- Quality: according to Laminar’s reported figures, flow-1 matches GPT-6-sol in failure detection on a benchmark of 523 difficult agent traces.
- Cost: flow-1 processes 23x as many traces per dollar as GPT-6-sol and is 25% cheaper than GPT-6-luna.
- Pricing: $0.05 per million input tokens, $0.01 for cached input and $0.30 per million output tokens.
- How it works: it runs inside Laminar’s Signals agent, which treats a trace like a code repo and each step like a file.
- Access: Laminar is open source under Apache 2.0 and can be self-hosted with Docker Compose; the company earns money from its managed cloud.
Watch: our 30-second video summary (in Vietnamese).
What Is Laminar flow-1?
An AI agent is a program that uses a language model to complete multi-step tasks, such as editing a contract or answering a customer. A trace is the full record of one agent run: every model call, every tool call and every result. Each of those steps is called a span.
The problem, as Laminar puts it, is that agents “fail implicitly.” An agent can click the wrong button or cite the wrong document while every single call completes successfully. Nothing crashes, so standard monitoring stays green. Looking at the final answer alone is not enough to catch it.
Laminar flow-1 is a small open-source model that Laminar fine-tuned and then trained with reinforcement learning (RL), a method where the model improves by being rewarded for good results. It was trained on thousands of synthetic investigations of realistic agent workflows in software, legal work, customer support and more. The goal: a model that is good at trace analysis and cheap enough to run at scale.
How Does It Find Errors in Agent Traces?
Laminar flow-1 does not read a trace top to bottom. It runs inside a harness called Signals, which works like a coding agent:
- It starts with a compressed preview of the full trace and a user prompt, such as “identify any logical errors.”
- It treats the trace as a repo and each span as a file.
- It greps across spans and opens the relevant ones in full to gather evidence.
- When it finds an issue, it returns structured JSON in a schema you define. Laminar calls these results signal events.
During RL training, the model used the same tools it has in production. It learned to decide what to inspect, collect evidence and judge whether that evidence really supports a finding.
Signals is not limited to bugs. You can ask it to spot user frustration, check whether a browser agent completed a checkout, or find out which requests agents hand off and why. Similar events are grouped into clusters, so you can see recurring problems and how often they happen. Traces from a cluster can then become an eval dataset to test a fix before release.
Real errors it caught
Laminar published example runs. In one, a legal agent marked a 105-day film availability window as agreed, but left the contract schedule at 90 days. flow-1 caught the error and linked it to the misleading “conformed clean” status in the issues list.
In another, a coding agent said its work was complete and blamed failing tests on a pre-existing issue. flow-1 noticed that the explanation only covered one of the failures. GPT-6-luna found no error in that trace.
Laminar is also open about gaps. In one contract review, flow-1 missed mandatory AI-governance protections that both GPT-6 models caught.
Laminar flow-1 vs GPT-6-sol: Benchmarks
Laminar ran the same Signals agent with different models on 523 difficult traces across coding, legal, customer support, CRM and other areas. These are the company’s own results.
| Model | Detection quality | Cost per trace (relative) | Notes from Laminar |
|---|---|---|---|
| Laminar flow-1 | Matches or beats GPT-6-sol | 1x (cheapest) | Digs into evidence, less noise |
| GPT-6-sol (high reasoning) | Strong | About 23x flow-1 | Flags more traces, noisier |
| GPT-6-luna | Weaker | About 1.33x flow-1 | flow-1 is 25% cheaper |
Laminar says GPT-6-sol flags more traces, which makes it noisier, while flow-1 focuses on the failures that matter.
Laminar flow-1 Pricing
| Token type | Price per million tokens |
|---|---|
| Input | $0.05 |
| Cached input | $0.01 |
| Output | $0.30 |
The cost of analysis includes the whole Signals agent run, not just one model call. Laminar’s comparison uses traces with under 100k total LLM tokens. The company says another RL checkpoint is training to make flow-1 a more efficient thinker and lower costs further.
How to Self-Host Laminar
The Laminar platform is Apache 2.0 licensed. The quickest way to try it locally, according to the official README, is Docker Compose:
git clone https://github.com/lmnr-ai/lmnr
cd lmnr
docker compose up -d
The UI then runs at http://localhost:5667. You also need to point the Laminar SDK at your instance with the right baseUrl and ports. For production, Laminar recommends its managed platform or docker compose -f docker-compose-full.yml up -d.
To send traces, add the Laminar SDK to your agent. It works with the Claude Agent SDK, OpenAI Agents SDK, AI SDK, LangChain, Mastra, Stagehand and more. Then describe what to look for in plain English, and Signals analyzes each trace.
One caution: the sources do not say whether Laminar flow-1 itself runs in a self-hosted setup. The managed cloud at laminar.sh is the documented way to “try flow-1 in Signals.” Check the docs before you plan around a fully local setup.
If you build agents with coding tools, see our guides on Claude Code subagents, Cursor rules for agentic workflows and Playwright MCP browser automation.
Is It Worth It for Your AI Agents?
Pros
- Very low price makes it possible to analyze every trace, not a sample.
- Catches “silent” failures where every call succeeds but the result is wrong.
- Flexible: any plain-English prompt and any JSON schema.
- Open-source platform you can self-host.
Cons
- All benchmark numbers come from Laminar itself.
- It still misses some errors that GPT-6 models catch.
- It is most useful for teams already running agents in production.
Laminar flow-1 FAQ
What does Laminar flow-1 do?
It reads AI agent traces and reports hidden failures, such as wrong actions or misleading status reports, as structured JSON.
How much does Laminar flow-1 cost?
$0.05 per million input tokens, $0.01 for cached input and $0.30 per million output tokens.
Is Laminar flow-1 better than GPT-6-sol?
On Laminar’s own benchmark, it matches or slightly beats GPT-6-sol in detection quality at 23x lower cost.
Is Laminar open source?
Yes. The Laminar platform is Apache 2.0 and self-hostable via Docker Compose.
Which frameworks work with Laminar?
The SDK supports the Claude Agent SDK, OpenAI Agents SDK, AI SDK, LangChain, Mastra, Stagehand and more.
Sources
- Launch HN: Laminar (YC S24) – RL-trained model that finds errors in agent traces
- Laminar blog: flow-1
- Laminar blog: flow-1 in action
- Laminar on GitHub
Last updated: October 20, 2026