Vercel CEO Guillermo Rauch
Guillermo Rauch describes several changes in Workflow SDK 5 and Vercel Workflows. Vercel says it reduced both the wait before the first step and the wait between steps. The team also strengthened encryption and tested scenarios including disaster recovery at scale. The underlying execution log is reusable, so Vercel packaged the durable execution mechanism as durable-harness for existing agent harnesses.
Garry Tan, President and CEO of Y Combinator
Garry Tan shares an open-source repository for evaluating agent harnesses. He compared Fable 5.2 and Opus 5.5 on 52 coding tasks in Claude Code 2.1.322. In his reported results, Fable solved 82.7% and Opus solved 84.6%, while Fable cost 2.80 times as much per solved task. On that basis he recommends keeping Opus 5.5 as the coding default in that environment. Opus already solves most tasks in the set, so Tan says harder tasks are needed to distinguish the two.
FPV Ventures partner Nikunj Kothari
Nikunj Kothari argues for trying models on your own workflows. He says that tuning for a single leaderboard can produce effective marketing, while leaving performance on another workflow unanswered. He sees value in a small number of general benchmarks. His post recommends choosing models against the work at hand and offers no comparative experiment.
Training Data — Ronak Malde on Training Models from Real Tasks
Key takeaway: Trajectory co-founder Ronak Malde wants to turn tasks users have actually performed into training material, then test improvements on similar tasks, deploy the model, and collect the next round of feedback.
At Windsurf, Malde used in-product edits and retries to train models. He compares today's batch updates to getting a paper route map before driving. The goal is closer to a navigation system that keeps learning during the trip. For now, the team brings custom model training, evaluation, and deployment into one workflow so companies that already have agents can put model updates into production. It's a step toward continually learning online.
He treats the complete task as the unit of work. To turn past work into training experience, first assemble the main agent call, subagent calls, product events, and user feedback into one trace. Then find a way to replay it and ask domain experts to define a good outcome. Recreating the real environment can be difficult, and legal or investment banking workflows may have no readily verifiable reward. Those are the two places where he wants to reduce manual work.
One research direction modifies GRPO so that training can learn from a task performed once, without repeatedly sampling the same environment. He also gives an example of selecting 100 representative traces out of 10,000, having experts write detailed criteria, and training a grader from those samples to evaluate more traces. He presents these as research directions and gives no assurance that production feedback can serve directly as a reliable score.
On Harvey's legal agent benchmark, he says the customized model raised accuracy from an open-source baseline of about 4% to about 18%. He also says it cost about one-sixth as much as frontier models. These are guest-reported results for a specific case and cannot be generalized to other tasks. For agents that might eventually explore on their own, he wants people to retain control over the goals and the system to learn from the results of interacting with users.