Hill-Climbing Skills: Browserbase Engineer Shows Agents Improving Without Touching Model Weights
AI Engineer · youtube · 2026-10-12
At AI Engineer World's Fair, Browserbase's Shubhankar Srivastava demonstrated how browser agents can improve at tasks without changing model weights.
Key points:
- Doer + teacher loop (AutoBrowse): a doer operates the browser while a teacher reads screenshots, traces and results after each run, writing lessons into a strategy file so the next attempt starts from accumulated experience
- Live demo: on a Google Flights task, a failed run discovered a URL shortcut that encodes the search directly, avoiding repeated calendar clicks; earlier conference-schedule and restaurant examples show skill files preserving navigation choices, waiting behavior and site-specific constraints, with replays making results inspectable
- Portability requirements: a replayable environment and verifiable outcome; discussion covers task-specific vs general skills, optimization objectives, iteration costs and verifying the verifier
- Ops & security: monitoring plus failure-triggered repair loops for changing websites, with human review; generated skills need scrutiny against prompt injection, alongside controls on inputs, allowed actions and destinations
Code is open-sourced at github.com/shubh24/ai-engineer-autobrowse.
More from coding & agent
- banteg builds a DSL to describe all 50 original game quests — banteg · 2026-10-12
- Bittensor SN33 claims skills turn small models into frontier-level performers — markjeffrey · 2026-10-12
- Decision 3.0 ported to Core ML: 1,000 PRs triaged on-device at 0.12s each — emax · 2026-10-12
- Dev claims open-source Schematic Tracer beats Cadence and Synopsys, built in 3 days — dosco · 2026-10-12
- mitsuhiko floats folding effect into code mode as durable execution wins praise — mitsuhiko · 2026-10-12
- Buddy 2.1: open-source Android app lets Claude Code drive your phone on your subscription — 3amae404 · 2026-10-12