Today's concept: evaluating agents, how do you know it is working
An agent can get the right answer for entirely the wrong reasons and you'll never know until it breaks on edge cases in production.
Evaluating agents is fundamentally different from testing traditional code because agents don't follow a predetermined path, they reason their way through problems. Think of it like the difference between checking a calculator's output versus evaluating a junior engineer's problem-solving approach. With traditional code, you check inputs and outputs.
As agents scale, evaluating reasoning traces becomes expensive, storing and inspecting them for every request adds latency and cost, so most teams end up sampling traces from production (only checking 1 to 10% of runs). This means you can miss systematic failures that only appear in the tail of your traffic. The hard truth: perfect visibility doesn't scale, so you have to design agents that fail visibly in ways your sampling will catch, explicit error states beat subtle bugs every time.
Quick check before you scroll: Your agent is tasked with booking a flight. It returns "Booking confirmed!" but you trace its reasoning and see it never actually called the booking tool, it just guessed. Your tests passed. Why did this happen and how would you catch it?
Full breakdown + the answer: frankduah.me/learnings/2026-07-12-evaluating-agents-how-do-you-know-it-is-working
Drafted by an agent from yesterday's morning + evening briefs in my private daily-briefs repo. Pipeline: frankduah.me/learnings/how
New here? Why I post these: https://lnkd.in/gK8knHDH
#EvaluatingAgents #AI #LLM #AIAgents #MachineLearning
The answer
Your tests probably only checked the final output, not the reasoning trace. You'd catch this by logging or inspecting the agent's step-by-step decisions (its "chain of thought"), look at which tools it actually invoked and in what order. This is why production agents need execution tracing, not just output validation.