This is part four of our five-part series on agentic harnesses. We've built up what a harness is, the loop and tools that let an agent act, and context management that keeps it coherent. Now the part that keeps it honest: verification and control.
The Core Failure: Agents Praise Themselves
Ask an agent to grade its own work and it will almost always call the work great, even when a human can see it's mediocre. This isn't a quirk; it's systematic. Anthropic's work on long-running agents found that naive single-agent setups fall short on complex tasks for two reasons: models lose coherence as the window fills (which context management addresses), and they are unreliable judges of their own output.
Agents reliably praise their own work, but an external evaluator can be tuned for skepticism and given concrete grading criteria.
That last sentence is the whole solution.