COLM hosts Self-Improving Agents social with talks on Meta-Harness, EvoSkill and agent fragility
tuvllms · x · 2026-10-06
A Self-Improving Agents social at COLM will feature four talks and a panel. The framing: self-improving agents split into two approaches—changing model weights, or keeping weights frozen and improving the surrounding system (prompts, tools, skills, memory, harness). Both face the same core problem: making a change is cheap, but knowing whether it actually improved the agent is much harder.
Talks:
- Seth Karten (Prime Intellect): Prime Agent and the Continual Harness—test-time scaling via hierarchical computation and memory, with reset-free refinement and trace hillclimbing
- Yoonho Lee (Stanford): Meta-Harness (COLM 2026) and follow-ups
- Qinyuan Ye (Salesforce): On the Fragility of Self-Improving Agents
- Noah Provenzano (Virginia Tech): EvoSkill and skill evolution
Panelists include Huaxiu Yao (Agent0) and Zora Wang (Agent Workflow Memory, SkillWeaver).
More from coding & agent
- Stack Overflow 2026 survey: 66% use AI coding tools, only 24% of firms have AI policies — pchandrasekar · 2026-10-06
- Shaders open-sources 200+ WebGPU effect components with free code export for React, Vue, Svelte, Solid — nicolascraske · 2026-10-06
- SOTA TV Podcast Launches With OpenHands Co-Founder and CMU Prof Graham Neubig — Madisonkanna · 2026-10-06
- DHH: To enjoy building with agents, you have to actually care about the software — csuwildcat · 2026-10-06
- Heavy Claude User on Gemini 4 Argon: Doesn't Beat Claude for Code, Antigravity Feels Alien — MicahBerkley · 2026-10-06
- ChatGPT Agent Answers Real Phone Calls With a $6 ESP32, Undercutting BPOs at $32/Month — jxnlco · 2026-10-06