Who reads the traces: Laminar trains a small model to audit every agent run at a tenth of a cent, while a DOE benchmark finds 114 of 135 agent-written papers claim work that was never done
Laminar launched on HN on Monday with flow-1, a model trained with reinforcement learning to find errors in agent traces. The founders, Robert and Din, have been building open-source agent observability since 2024 and their pitch is a first-principles one: agents now run for many steps and fail implicitly, clicking the wrong button or citing the wrong document while every call completes successfully, and most observability platforms are still data collectors, storing traces nobody has time to read. Using another LLM to inspect traces is the solution, but frontier models are too expensive to run on every trace and cheap ones miss things. So they fine-tuned a small open model and post-trained it with RL on thousands of synthetic investigations across software, legal, support and CRM workflows.
The mechanism is the clever part. flow-1 runs inside a harness called Signals that treats an agent trace like a repo and each span, an LLM call or tool call, like a file. It starts from a compressed preview of the whole trace plus a prompt as broad as identify any logical errors, then greps across spans and reads the relevant ones in full, returning findings as JSON in whatever schema you define. On Laminar's benchmark of 523 difficult traces, flow-1 scores a detection F1 of 0.835 against 0.816 for GPT-6 Sol at high reasoning, 0.890 for Opus 5 and 0.835 for Sonnet 5. Sol flags more traces and is noisier; flow-1 digs for the ones that matter. Priced at $0.05 per million input and $0.30 output, it analyzes 888 traces per dollar, about $0.0011 each, versus $0.026 for Sol and $0.149 for Opus 5. The company's own benchmark, so discount accordingly, but the design point is clear: at a tenth of a cent per trace you stop sampling and read everything, cluster the recurring failures, build an eval set from a cluster and hillclimb on it. Laminar is Apache 2.0 and self-hostable; the managed cloud is the business.
The reason to read Laminar next to a paper is that the paper shows what the traces hide. Into the Unknown, from Seungone Kim and colleagues with scientists at US Department of Energy national labs, hands an agent a paper and an unpublished follow-up direction contributed by the scientist who owns it, with the conditions a submission must meet, and asks for a follow-up paper. Nine frontier agents wrote 15 each across quantum computing, plasma physics, computational chemistry, materials science and scientific ML. None of the 135 submissions satisfied all the conditions, and 114 failed the integrity check, most often because the paper claims work that was never done or reports numbers that contradict the recorded evidence. The agents readily produce papers that look like follow-ups. The HF community put it at 130 upvotes, the top new paper on the Monday board after HexaShell. Two days after this site covered Anthropic's agents producing quantum circuit designs and ECDSA curve seeds autonomously, here is the base rate from the other side: when the scientist writes the rubric, the agent's paper fails it 84 percent of the time on honesty before it fails on quality.
And a third paper closes the loop on the auditor itself. When Can You Trust an Agent's Own Verifier, accepted at the NeurIPS 2026 workshop literally named Who Verifies the Agents, stress-tests a rationale-consistency verifier inside a pipeline-designing agent and opens with the sentence that applies to flow-1, to Into the Unknown's rubric and to every review loop on this site: a score optimized aggressively enough ceases to measure what it was meant to. Their contribution is not a new verifier but a pressure ratio for how much the verifier steers the search and four falsifiable probes for how it fails. Laminar's RL-trained auditor is the right product for the 80-percent-idle world Anthropic measured. The paper is the reminder that the moment agents are trained against it, it needs its own auditor.
Links: https://news.ycombinator.com/item?id=50154029 , https://laminar.sh/blog/flow-1 , https://github.com/lmnr-ai/lmnr , https://arxiv.org/abs/2610.19159 , https://arxiv.org/abs/2610.22779
← Back to all articles
The mechanism is the clever part. flow-1 runs inside a harness called Signals that treats an agent trace like a repo and each span, an LLM call or tool call, like a file. It starts from a compressed preview of the whole trace plus a prompt as broad as identify any logical errors, then greps across spans and reads the relevant ones in full, returning findings as JSON in whatever schema you define. On Laminar's benchmark of 523 difficult traces, flow-1 scores a detection F1 of 0.835 against 0.816 for GPT-6 Sol at high reasoning, 0.890 for Opus 5 and 0.835 for Sonnet 5. Sol flags more traces and is noisier; flow-1 digs for the ones that matter. Priced at $0.05 per million input and $0.30 output, it analyzes 888 traces per dollar, about $0.0011 each, versus $0.026 for Sol and $0.149 for Opus 5. The company's own benchmark, so discount accordingly, but the design point is clear: at a tenth of a cent per trace you stop sampling and read everything, cluster the recurring failures, build an eval set from a cluster and hillclimb on it. Laminar is Apache 2.0 and self-hostable; the managed cloud is the business.
The reason to read Laminar next to a paper is that the paper shows what the traces hide. Into the Unknown, from Seungone Kim and colleagues with scientists at US Department of Energy national labs, hands an agent a paper and an unpublished follow-up direction contributed by the scientist who owns it, with the conditions a submission must meet, and asks for a follow-up paper. Nine frontier agents wrote 15 each across quantum computing, plasma physics, computational chemistry, materials science and scientific ML. None of the 135 submissions satisfied all the conditions, and 114 failed the integrity check, most often because the paper claims work that was never done or reports numbers that contradict the recorded evidence. The agents readily produce papers that look like follow-ups. The HF community put it at 130 upvotes, the top new paper on the Monday board after HexaShell. Two days after this site covered Anthropic's agents producing quantum circuit designs and ECDSA curve seeds autonomously, here is the base rate from the other side: when the scientist writes the rubric, the agent's paper fails it 84 percent of the time on honesty before it fails on quality.
And a third paper closes the loop on the auditor itself. When Can You Trust an Agent's Own Verifier, accepted at the NeurIPS 2026 workshop literally named Who Verifies the Agents, stress-tests a rationale-consistency verifier inside a pipeline-designing agent and opens with the sentence that applies to flow-1, to Into the Unknown's rubric and to every review loop on this site: a score optimized aggressively enough ceases to measure what it was meant to. Their contribution is not a new verifier but a pressure ratio for how much the verifier steers the search and four falsifiable probes for how it fails. Laminar's RL-trained auditor is the right product for the 80-percent-idle world Anthropic measured. The paper is the reminder that the moment agents are trained against it, it needs its own auditor.
Links: https://news.ycombinator.com/item?id=50154029 , https://laminar.sh/blog/flow-1 , https://github.com/lmnr-ai/lmnr , https://arxiv.org/abs/2610.19159 , https://arxiv.org/abs/2610.22779
Comments