You've Been a Bad Agent Matt Goes Viral with Server-Side Code Mode, Wilhelm Loses His Files, and What Even Is a Sandbox?
Feb 23, 2026
Surfing life updates from Lisbon set a relaxed tone. A viral thread on moving code execution to server-side Code Mode sparks debate about MCP and simplified client implementations. Deep dives into agent-first tooling, multi-model review workflows, and sandbox types from VMs to dynamic isolates. A near-disaster rsync story leads into improved backups, sync tools, and automated healthcheck monitoring.
AI Snips
Chapters
Transcript
Episode notes
Prefer Minimal Harnesses For Modern Models
- Avoid over-building harnesses for models; simpler tool sets often suffice as models improve.
- Wilhelm notes Pi's tiny toolset (read/write file, bash) already enables many agent tasks because models are stronger now.
Stronger Models Need Less Tool Chatter
- Newer Codex models are less trigger-happy and can produce more complete solutions when given clearer specs.
- That shifts agent design from frequent tool calls to letting models think through edge cases and return finished results.
Claude And Codex Bounce Ideas Together
- Matt runs a workflow where Claude and Codex 'bounce ideas' by calling each other via CLI sessions for planning.
- He has a skill that instructs Claude to call another model with a session ID and then they chat back and forth.