COLM Social: Self-Improving Agents
Agents that improve themselves are showing up everywhere in this year's program: agents that write their own skills, loops that rewrite the harness, models trained on their own verified outputs. One approach changes the model weights. The other keeps the weights frozen and improves the system around the model: prompts, tools, skills, memory and the harness. Both face the same problem: making a change is cheap, knowing whether it actually made the agent better is much harder.
Agenda
1:00 PM · Welcome
1:10 PM · Talk · Seth Karten
1:20 PM · Talk · Yoonho Lee
1:30 PM · Talk · Qinyuan Ye
1:40 PM · Talk · Noah Provenzano
1:50 PM · Panel and audience Q&A
2:25 PM · Wrap-up
Talks
Seth Karten · Research at Prime Intellect, PhD candidate at Princeton.
Talk: Self-Improving Agents through Prime Agent and the Continual Harness. Prime Agent is an RLM harness for test-time scaling through hierarchical computation and memory, extended by the Continual Harness with reset-free refinement, software factory abstractions, and trace hillclimbingYoonho Lee · PhD student at Stanford. First author of Meta-Harness (COLM 2026)
Talk: Meta-Harness and more recent follow-ups.Qinyuan Ye · Research Scientist at Salesforce AI Research. First author of On the Fragility of Self-Improving Agents.
Noah Provenzano · PhD student at Virginia Tech
Talk: EvoSkill and skill evolution.
Panel
Huaxiu Yao · Assistant Professor at UNC Chapel Hill. Co-author of Agent0.
Zora Wang · PhD student at Carnegie Mellon. Agent Workflow Memory, SkillWeaver.
Chien-Sheng (Jason) Wu · Senior Director at Salesforce AI Research.
Moderated by Raphael Kalandadze.
Organizers
Good to know
For registered COLM 2026 attendees: you need your conference badge to get in.
It's during the lunch break, so feel free to bring your lunch.