Debug log: harnesses fail like Heisenbugs

By Rohan Kataria · Mar 25, 2026

Key points

  • The expectation is that a harness isolates agent state, so failures are 100% reproducible.

  • In production, context leakage, ephemeral file handles and ambient tool state cause non-deterministic failures.

  • The fix is to treat each harness as the OS, with immutable snapshot boots, explicit tool scoping, replayable timelines and reasoning-level early warning.

  • The harness isolates the environment, and monitoring catches semantic displacement.

Updates and sources

  • Immutable snapshot boots start every agent from a clean, frozen state with no residue.

  • Explicit tool scoping removes ambient access. If a tool isn't in the MCP definition, it doesn't exist.

  • Replayable timelines rehydrate the exact state to debug a failure, instead of chasing a static log.

  • Reasoning-level early warning monitors semantic displacement between steps, as described in the ProMAS paper, and catches failures before the cascade.

Sources I cited in the comments: