Harrison Chase 谈 harness 时把 Jev 列为决策模型范例
LangChain 创始人到访 Profound 谈 agent harness 构建，提到接入 Jev 这类更快更便宜的决策模型是通用 harness 的优势示例。
The 🐐 @hwchase17 paid a visit to @profound HQ in NYC today to talk about building harnesses at the frontier. Takeaways: 1/ He has one of the sharpest definitions of what a harness is: the orchestration logic that gets the right context to the model at the right time. 2/ An off-the-shelf harness is better to start with compared to a model-specific SDK, as it has both extensibility and model flexibility. Combining access to faster and cheaper decision models like Jev is a great example. 3/ A context layer that only ingests sources into a structured set of context files is no longer enough. A great harness needs a context layer that relies on "online learning," where agents review interactions, surface gaps, and suggest updates continuously. 4/ Developing great Evals is essential for 1) defining a benchmark for what "good" looks like internally and 2) enabling hill-climbing to improve the harness. 5/ Agent tracing and observability close the loop. By evaluating each step in an agent run, teams can tell whether a harness is improving or adding friction. Awesome session, thank you Harrison for coming in!