AI dev stack event in San Francisco focuses on production-ready agents

The second event in the The AI Dev Stack series targets the layer beneath code agents, the infrastructure that keeps them running in production. Antimetal, Fireworks, Braintrust and Browserbase will each give a seven-minute demo followed by a Q&A. The gathering is a focused meetup for developers who build these layers daily, not a large conference.

Antimetal: Shreyas Iyer, CTO and co-founder, will present Triage, a tool that clusters events related to issues and learns which signals matter to the team. The premise is simple: instead of flooding developers with thousands of alerts, the system identifies recurring patterns and surfaces only what requires intervention. Iyer did not publish performance metrics or comparative data with existing solutions, so the effectiveness at scale remains unclear.

Fireworks: Sinan Ozdemir, head of AI developer education, will demonstrate turning failed evaluation cases into fine-tuning data. The methodology compares three tracks: a frontier model API, a base-version open model, and the same open model after fine-tuning on its own failures. The approach closes the loop between evaluation and training, but without public benchmarks it is hard to assess whether the results exceed a commercial API.

Braintrust: Hossein Niazmandi, head of solution engineering, will address the familiar problem of an agent that works in demo but fails when real users arrive. Their solution converts real production failures into evaluations that capture regressions before they reach customers. It is essentially CI/CD for agent behavior; the architectural idea is sound, yet the implementation depends on instrumentation quality and the ability to reproduce failures in a development environment.

Browserbase: Shrey Pandya, growth engineer, will show how the platform gives agents the “fleet”, tools and access they need to navigate the web better than a human. The practical output includes a managed browser with automation, CAPTCHA handling, session and identity management—everything an agent needs to operate on sites not built for APIs. Browserbase did not disclose comparative performance data against Playwright or Puppeteer in an independent setup.

What emerges from the line-up is a consensus shift: the question is no longer “how to write code with LLMs” but “how to monitor, evaluate, fix and grant external-world access”. Each company attacks a different angle—alert noise, evaluation-training loop, production reliability, web access—and none presents a standalone “code agent”. That stack must be built before agents become stable products.