Article
Laminar brings Agent the cost of error investigation down to GPT-6-sol of 1/23 Beating AI news,AI Agent observability platform Laminar has released a model specifically for checking Agent execution traces. It reads model calls, tool calls, and returned results to identify flow-1 at which step the error occurred and why it occurred.Agent flow-1 It runs in Laminar's Signals agent. The model can search the entire trace（Agent(the complete execution record), then inspect specific steps as needed. During training, Laminar first used synthetic investigation data for supervised fine-tuning, then used reinforcement learning to train tool calling and complex trace analysis.
In Laminar's self-built 523 difficult trace benchmark,flow-1 error detection F1 was 0.835，GPT-6-sol was 0.816 sol has higher recall and can find more real errors;flow-1 has higher precision and fewer false positives.
For 100000 LLM tokens of fewer than trace，flow-1 the average analysis costs about 0.0011 USD，GPT-6-sol about 0.026 USD According to Laminar's estimates, 1 USD can analyze about 888 and 38 respectively, a difference of about trace times. 23 The cost of flow-1 is also about GPT-6-luna lower than 25%.
At present, these results all come from Laminar's self-built benchmark, with no third-party retesting to date.flow-1 training data mainly comes from synthetic workflows, of which about 48%consists of software engineering tasks. Some in the community have already questioned whether it can maintain the same performance when facing novel failures in real production environments that were not designed in advance.
Original link https://m.theblockbeats.info/flash/370475