Laminar Launches flow-1 to Read Every Agent Trace for Errors

Laminar, a six-person Y Combinator Summer 2024 company in London, launched flow-1 on Launch HN, describing it as a small open-source model trained with reinforcement learning to find errors in AI agent traces. The founders claim it matches a frontier model's detection quality at about 23 times lower cost. That figure is vendor-supplied, and the post names neither the frontier model nor a test set.

The post stood at 59 points on Oct. 24. It is signed by Robert Kim, chief executive, and Din Mailibay, chief technology officer, per Y Combinator's company page.

How flow-1 works

Laminar's harness, called Signals, treats an agent trace like a code repository, with each span as a file. flow-1 greps and reads those spans and returns findings as JSON in a schema the customer defines. Similar findings are clustered so recurring failures and their frequency show up, and clusters can become evaluation datasets for testing a fix before release.

The cost claim is the commercial point. At 23 times cheaper, the founders argue, a team can analyse every production trace instead of sampling a few. The post does not say whether the model weights are downloadable; it calls flow-1 open source, and The Terminal did not check a model repository.

What the repo shows

The lmnr repository is Apache-2.0 with about 3.4k stars, 249 forks, 65 open issues and 77 open pull requests, as of Oct. 24. The core is written in Rust, with Python and TypeScript SDKs. Self-hosting runs through docker compose up -d for a lightweight stack; the docs steer production use to the managed cloud or a full compose file backed by ClickHouse. Stars measure attention, not adoption.

Item	Detail
Licence	Apache-2.0
Claimed trace compression	About 20x (vendor-supplied)
flow-1 cost claim	About 23x below a frontier model (vendor-supplied)
Pricing	Not published in the post or repo

Money and customers

Laminar raised a $3 million seed on March 19, 2026, Pulse 2.0 reported, led by Atlantic with Y Combinator and AAL VC participating. Kim named Browser Use, OpenHands and Rye as customers. The founders say revenue comes from the managed cloud, but give no pricing or revenue figure, so no multiple can be computed.

1. Unnamed frontier model23 x flow-1 cost
2. flow-11 x flow-1 cost

Source: Launch HN post by Laminar, accessed 2026-10-24 (vendor-supplied)

What commenters asked

Hacker News readers went straight at calibration. User ianalloway1 asked how the model is checked against human labels when judge models can share an agent's blind spots, and khalidsaidi said many failures in their transcripts were environment faults such as 403 errors, not reasoning errors, and asked whether flow-1 separates them. The founder replied that training targets this. User pmichalas asked for calibrated probabilities instead of severity labels. The thread offers no false-positive rate.

Laminar sits among agent tooling such as the open-source HumanLayer, and our coverage of OpenAI's offline training shows why labs want to inspect agent behaviour closely. The next thing to look for is a model card or benchmark for flow-1, which the post does not promise.