Cursor How Two Baseten Researchers Use Coding Agents
27 snips
Jun 11, 2026 
Harry Partridge, an engineer building large-scale agent tooling at Parsed, and Charlie O'Neill, a researcher specializing in open-source model specialization, discuss running dozens to hundreds of coding agents. They cover agent coordination, agent-to-agent messaging, automated QA and adversarial code reviews, mixing models for implementation versus critique, and reusable skills and workflows for scaling agent systems.
AI Snips
Chapters
Transcript
Episode notes
Running Hundreds Of Specialized Agents
- Charlie runs 64–128 agents for KVCache compaction research and usually has 16 actively running jobs across many GPU nodes.
- He delegates tasks to named subagents (e.g., Hilbert, Poincaré) and monitors them via a CLI view showing many agents' messages.
Use Judge Agents To Prevent Premature Stopping
- Use separate judge agents to verify completion so a main agent can't prematurely claim success.
- Build loops and reminders that force agents to continue until an independent verifier confirms the task is done.
Taste Remains The Bottleneck For Nonverifiable Tasks
- Agents are good at writing code but still weak at reading and forming global understanding; human taste is the bottleneck for non-verifiable tasks.
- Charlie and Harry inject 'taste' and steer models when tasks need deep causal thinking or aesthetics.