Hill-Climbing Skills: Improve Agents Without Changing the Model — Shubhankar Srivastava, Browserbase
The talk introduces 'hill-climbing skills,' a methodology to significantly improve the reliability and performance of browser agents without requiring changes to the underlying large language model weights. The core concept is the AutoBrowse loop, which pairs a 'Doer' agent (executing browser actions) with a 'Teacher' agent (analyzing traces, screenshots, and logs). The Teacher learns from the Doer's failures and successes, writing these 'lessons' into a persistent `strategy.md` file. This artifact guides subsequent runs, allowing agents to discover complex, site-specific hacks (like using query parameters instead of clicking through forms) and dramatically reducing cost and improving success rates in non-deterministic environments.
Key takeaways
-
The Doer-Teacher Architecture
15:32
The AutoBrowse loop consists of two agents: the Doer, which performs the browsing actions using the `browse` CLI, and the Teacher, which analyzes the Doer's traces and identifies improvements. This collaborative process continues until a successful, converged solution is found. (9:32)
-
Skill Persistence via Strategy Files
15:32
Learnings are persisted in a `strategy.md` file, which acts as a reusable artifact for future agents. This file contains hacks, constraints, and optimal navigation choices, preventing the agent from 'relearning' tasks from scratch. (9:32)
-
Addressing Reliability and Cost
18:50
By automating the discovery of optimal paths (e.g., using a direct URL query parameter instead of clicking through a calendar widget), the system achieves massive savings in cost and turns, demonstrating improvements of up to 16x cheaper on different tasks. (11:30)
-
Verifiable Outcomes are Critical
29:10
For the AutoResearch loop to function, two conditions are required: a replayable environment and a verifiable, quantifiable outcome (metric). The difficulty lies in defining a metric when the goal is subjective (e.g., finding the 'best' flight vs. finding 'a' flight). (17:50)