# Eval Quality Is the Rate-Limiter on Safe Agent Autonomy A coding agent can only safely self-improve as fast as its evals give it a clean feedback signal, so the quality of the evaluation — not the model — sets the ceiling on how much autonomy you can hand it. ## Scope & conditions - Holds when an agent loops on its own output: the eval is the signal it optimizes against, so a noisy or gameable eval caps how far you can let the loop run unattended. - "Quality" means the eval actually measures what good looks like and is stable enough to trust as a gate; a flaky eval is worse than none because the stop condition reads its verdict. - Breaks down where outcomes can't be cheaply scored (open-ended judgment, taste, novel domains); there autonomy stays low regardless of model capability. ## Links - [[3. Resources/Permanent Notes/Encoding What Good Looks Like Is What Makes Agent Verification Trustworthy]] — because extends; a checked-in standard of "good" is what an eval operationalizes, making the agent's self-review trustworthy - [[3. Resources/Permanent Notes/A Nondeterministic Check Breaks the Loops Stop Condition]] — because evidence-for; an unstable eval breaks the unattended loop even with an independent checker, since the stop condition depends on its signal - [[3. Resources/Permanent Notes/Bounded Action Spaces Not Better Models Make Agents Safe to Run]] — because contrasts on lever; that note bounds what the agent may *attempt*, this one bounds how fast it may *improve* — both make safety a property of the harness, not the model <!-- atomicity: 1 idea ✓ --> - [[3. Resources/MOCs/AI & the Future of Knowledge Work MOC]] — applies-in: the MOC lists this note under Amplifier thesis & the harness. - [[3. Resources/MOCs/AI Agent Safety MOC]] — applies-in: the MOC lists this note under Verification & the maker/checker split.