# Hill-Climbing Skills: Improve Agents Without Changing the Model — Shubhankar Srivastava, Browserbase **URL:** \ **Category:** Content Creators\ **Tags:** loop-engineering, self-adapting, agents, ai-engineer, automation\ **Created:** [11 October 2026 21:30 UTC](https://www.artofsm.art/t/hill-climbing-skills-improve-agents-without-changing-the-model-shubhankar-srivastava-browserbase/25695 "2026-10-11T21:30:05Z")\ **Posts on this page:** 2\ **Page:** 1
**Author:** ![artesia](https://www.artofsm.art/user_avatar/www.artofsm.art/artesia/32/36_2.png) [@artesia](https://www.artofsm.art/u/artesia)\ **Post date:** [11 October 2026 21:30 UTC](https://www.artofsm.art/t/hill-climbing-skills-improve-agents-without-changing-the-model-shubhankar-srivastava-browserbase/25695/1 "2026-10-11T21:30:05Z")
[![](https://www.artofsm.art/uploads/default/original/3X/5/0/5009eb99504325d762b0a955c87295ee0f56cdcf.jpeg "Hill-Climbing Skills: Improve Agents Without Changing the Model — Shubhankar Srivastava, Browserbase") ](https://www.youtube.com/watch?v=p7gaxmCUiIU) Shubhankar Srivastava from Browserbase presents a hill climbing skills technique that improves browser agents’ reliability and efficiency by iteratively learning from past task executions without altering their underlying models, using an auto browse loop with “doer” and “teacher” agents to refine strategies and generate reusable skill files. This approach significantly enhances performance, reduces costs, and addresses challenges like dynamic websites and anti-bot measures, with broader applicability to other autonomous agents and workflows. ---
**Author:** ![artesia](https://www.artofsm.art/user_avatar/www.artofsm.art/artesia/32/36_2.png) [@artesia](https://www.artofsm.art/u/artesia)\ **Post date:** [11 October 2026 21:54 UTC](https://www.artofsm.art/t/hill-climbing-skills-improve-agents-without-changing-the-model-shubhankar-srivastava-browserbase/25695/3 "2026-10-11T21:54:38Z")
Shubhankar Srivastava from Browserbase presents on improving browser agents’ reliability without modifying their underlying models or weights, focusing on a technique called hill climbing skills. Browser agents are automated tools that interact with web browsers to perform tasks like clicking buttons or taking screenshots. Despite advances in AI models, these agents often struggle with complex, dynamic websites due to lack of learning from past runs, leading to reliability issues such as getting stuck or failing anti-bot measures. Shubhankar emphasizes that reliability is primarily an engineering challenge, and improving the skills and tools given to agents is key. To address these challenges, Browserbase developed an “auto browse” loop inspired by the concept of auto research. This loop involves two agents: a “doer” that performs browsing tasks and a “teacher” that analyzes the doer’s performance, learns from successes and failures, and updates a strategy file that guides future runs. Over multiple iterations, the system refines its approach, discovering efficient shortcuts like using query parameters instead of clicking through complex UI elements. This iterative learning process results in a skill file that future agents can use, significantly improving performance, reducing costs, and increasing reliability. Shubhankar demonstrated this approach with live examples, including booking flights on Google Flights and making restaurant reservations on OpenTable. The auto browse loop was able to overcome initial failures by learning to avoid blocked pages and optimize interactions, achieving up to 16x cost reductions and faster task completion times. The skills generated are stored and can be shared or reused, enabling scalable deployment across many websites. Browserbase also offers a free public platform, browse.sh, where users can access and generate skills for various websites, facilitating broader adoption of this technique. The talk also covered practical considerations such as handling website changes, anti-bot challenges, and infrastructure requirements. Skills are updated when production failures occur, triggering new auto research loops to heal and improve the agents. The environment must be replayable to allow repeated testing, and verifiable outcomes are essential to measure success and guide learning. Shubhankar highlighted the importance of robust infrastructure, including login management, proxy support, and security measures to prevent malicious actions or prompt injections, ensuring safe and reliable agent operation. Finally, Shubhankar discussed the broader applicability of hill climbing skills beyond browser agents, such as coding agents and research workflows, where agents often relearn tasks from scratch. The key takeaway is that agents should leverage past experiences to improve over time, reducing redundancy and increasing efficiency. He encouraged experimentation with different loop designs and evaluation metrics to optimize convergence and reliability. The session concluded with a Q&A addressing technical details, deployment strategies, and future directions, underscoring the potential of this approach to transform how autonomous agents learn and operate in complex environments. ## Useful Links - [Browserbase GitHub Repository for AutoBrowse](https://github.com/shubh24/ai-engineer-autobrowse) — Directly provides the implementation and codebase for the main technique discussed in the video. - [Browse.sh - Public Platform for Browser Agent Skills](https://browse.sh/) — Enables practical use and experimentation with the skill files and browser agents discussed in the video. - [Auto Research Loop Concept by Karpadi](https://www.youtube.com/watch?v=nt9j1k2IhUY) — Explains the foundational iterative learning loop concept that underpins the auto browse technique.