Hill-Climbing Skills: Improve Agents Without Changing the Model — Shubhankar Srivastava, Browserbase

Shows how to improve a browser agent without touching weights by distilling run traces into a skill file. It also states the preconditions: a replayable environment and a verifiable outcome.

• Browserbase's AutoBrowse loop improves a browser agent without changing model weights: a doer operates the browser, and a teacher reads screenshots, traces and results after each run and writes lessons into a strategy file the next attempt starts from.
• In the Google Flights demo, a failed run discovered the search could be encoded directly in the URL, skipping repeated calendar clicks and a time filter. The strategy file then graduates into a reusable skill file for future agents.
• Shubhankar Srivastava says the approach transfers to domains with a replayable environment and a verifiable outcome, but notes the limit: judging an acceptable flight is easier than judging the best one, and generated skills need prompt-injection scrutiny.

• AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
• agent skill — A reusable instruction file that teaches an agent how to do one job well — the procedure, the tools, and what counts as done.

• videoHow Secrets Leak Through AI Agents (and How to Stop It) — Venice AI
• videoFrom Zero to Leaderboard: Agent Evaluation — Wolfram Ravenwolf, Weights & Biases
• videoBuild a Platform and Watch It Burn — Michael Forrester, Accenture & Whitney Lee, Datadog
• videoContinuously Improving Agents with Langfuse — Annabell Schäfer & Lotte Verheyden, ClickHouse

• videoLet Your Agent Cook: Using Skills to Evaluate and Improve Your App — Ankur Duggal, Arize AIAI Engineer
• videoBuilding Self-Learning Loops for Your Agent — Fuad Ali, Arize AIAI Engineer
• videoWhy 99% Accurate Browser Agents Still Fail — Derek Meegan, BrowserbaseAI Engineer

Checking sign-in…

Loading comments…