# Comprehension Debt Is the Real Cost of Unattended Autonomy ## Claim An unattended agent loop's sharpest risk is not token cost or error rate but comprehension debt — the gap that widens when a team ships code nobody actually read, so two engineers running the identical loop can land at opposite real-world outcomes depending on how much they understood of what shipped. <!-- atomicity: 1 idea ✓ --> ## Scope & conditions - Holds for any autonomous or scheduled loop producing shippable output (code, content, decisions) faster than a human reviews it. - The debt is invisible in the loop's own success metrics (tests passed, task marked done) because those measure the maker's claim, not the team's understanding. - Breaks where the loop's output is trivial, fully reversible, or independently re-derivable on demand — comprehension can be deferred without real risk there. ## Evidence The source frames "a loop running unattended" as "also a loop making mistakes unattended," and names the sharper risk beneath that: comprehension debt, the accumulating gap between what a system has shipped and what the team running it actually understands well enough to defend, extend, or debug. The same loop, run by two different engineers, can produce opposite real outcomes — not because the loop behaved differently, but because one engineer tracked what it did and the other didn't. ## Links - [[3. Resources/Permanent Notes/The Maker Should Not Grade Its Own Homework]] — extends: an independent checker catches whether the output is *correct*; comprehension debt is a separate failure mode where the output can be correct and still nobody on the team can explain or maintain it. - [[3. Resources/Permanent Notes/A Nondeterministic Check Breaks the Loops Stop Condition]] — applies-in: a determinate stop condition tells you the loop finished; it says nothing about whether a human understood what finished. - [[AI Agent Safety MOC]] — applies-in: comprehension debt is an unmonitored failure mode that belongs alongside verification and access-control risks this MOC tracks. - [[3. Resources/Permanent Notes/A Demo-Only Rubric Selects for Developers Who Cannot Explain What They Shipped]] — applied in: score the explanation as part of the rubric, not as a follow-up (Saha). ## Tests / falsifiers - Compare defect/incident rates on loop-shipped work between teams that require a human read-through before merge versus teams that don't, controlling for loop verifier quality. - Track whether comprehension debt correlates with loop run frequency (more runs, less time per human to absorb each one) or is independent of cadence. ## Maintenance - Last tended: 2026-08-16 - Next review: 2026-11-16 ## Source - [The Anthropic Leader Who Built Claude Code Says He Ditched Prompting — Now He Just Writes Loops.](3.%20Resources/Readwise/Articles/The%20Anthropic%20Leader%20Who%20Built%20Claude%20Code%20Says%20He%20Ditched%20Prompting%20—%20Now%20He%20Just%20Writes%20Loops..md) — names comprehension debt as the sharpest risk of autonomous loop-running, distinct from token cost or verifier quality.