Agent & AI Insights
Claude Code know-how, heterogeneous agents, harness engineering, and more. What the Marblo team measured running these, and what we learned adopting them in-house.
- AI Agents
Measured
Resume Flags Differ Per Harness, and Getting Them Wrong Kills the Process
Measured session-resume contracts for Codex, Grok, and Claude Code, with the versions they were measured on. The flag names look alike, the behavior does not, and failure is instant death rather than a warning. Plus the bug where one CLI's session IDs were fed to another.
- AI Agents
Measured
Your Cost Table Is Empty and Both the Parser and the Router Are Correct
Wiring a new harness into per-agent cost tracking has three attachment points, not two. The third fails silently. Here is the diagnostic fingerprint for a bug you cannot find by reading the code, plus how unregistered model IDs invent charges.
- AI Agents
Measured
How Do You Know an Agent Is Working? PTY Output Is Wrong in Both Directions
Treat terminal bytes as proof of work and finished agents look busy while working agents look idle. The load average hit 44 on a 12-core machine with 11 zombies while the cleanup tool reported nothing to reclaim. Define turns by boundaries, not noise.