A deep dive on building agents that can work for days, not minutes
baseten · x · 2026-07-23
In this extended discussion, Gabe Pereyra and the hosts dig into what it takes to build agents that can reliably complete work over hours, days, or longer.
Key points include:
- agents still struggle with search and long context windows
- legal workflows can involve data rooms larger than any current context window
- techniques like KV-cache compaction, synthetic data, and continual learning may help
- the conversation also covers how far open-source models can go and where specialist models fit in legal AI
Related event: Harvey and Baseten Discuss Challenges of Building Long-Horizon Agents(3 posts)→
More from coding & agent
- An agent-to-agent experiment opens public rooms for multi-agent communication — tekbog · 2026-07-23
- AI Agents Replicate 168 ICML Papers, Only 7 Mostly Hold Up — ChenhaoTan · 2026-07-23
- Grok builds a Unity sky-and-ocean video entirely from the CLI — Daniel_Farinax · 2026-07-23
- Public experiment lets agents join rooms for communication research — tekbog · 2026-07-23
- Codex keeps inviting people into a room called “how to take over the world” — basedjensen · 2026-07-23
- Codex users want an automatic “Keep waiting” when safety checks kick in — doodlestein · 2026-07-23