'Coherence Debt': Withheld Facts Make Coding Agents Fabricate Rather Than Stall, and Harnesses Differ Tenfold in Tokens for Identical Results
Across seven models and five harnesses, the authors supply and withhold context versus parametric memory and inject faults, finding availability — not distance from the edit — decides outcomes: a supplied fact works as well far from the edit as adjacent to it. Harness configurations that all pass every test differ more than tenfold in tokens consumed because they rebuild the same facts at different rates, and spending more recovers nothing when a fact is genuinely absent. Two findings sting for anyone instrumenting agents: a missing fact yields wrong work rather than absent work (the agent fabricates the file or guesses the value, so read-based instrumentation sees a hole already filled), and where a convention file and the code disagree, agents follow the stale convention — making a wrong CLAUDE.md worse than none.
↳ Follow the thread