AggAgent at COLM: Treats Parallel Agent Trajectories as an Environment for Long-Horizon Tasks
xiye_nlp · x · 2026-10-06
Researcher yoonsang is presenting the AggAgent poster at COLM (Tuesday 4:30–6:30pm, Franciscan B #88) and open to chats on agent training, harnesses, and test-time scaling.
Core idea: unlike CoT reasoning, agentic tasks are long, multi-turn, and tool-augmented. AggAgent addresses how to effectively aggregate long-horizon agent trajectories by treating parallel trajectories as an environment the agent can interact with.
More from coding & agent
- AgentHopper: a cross-agent 'AI virus' built on chained prompt injection — wunderwuzzi23 · 2026-10-06
- Perplexity Computer builds a flight simulator prototype for just 97 credits — AravSrinivas · 2026-10-06
- How often should you run evals? Hamel Husain lays out a 3-factor tradeoff framework — HamelHusain · 2026-10-06
- Gemini 3 Flash burned $14 and failed a $1 agent task Pro finished in 85 steps — XIFAQ · 2026-10-06
- This builder spends 5x more on scraping APIs than AI models — data is the edge — EXM7777 · 2026-10-06
- Interfaze open-weights MoA model for OCR, speech and GUI grounding on one 80GB GPU — charles_irl · 2026-10-06