Last week in AI evaluation and agentic AI
An agent that can do a task and an agent you can trust with it are two different claims, and a single benchmark score can blur them together.
On 19 October, researchers from Princeton, Stanford, the UK AI Security Institute and Brookings released "Emerging Trends in AI Agent Evaluation", a survey of 33 shifts in how agents get tested. The part I keep coming back to is how it treats repeated runs.
Two terms first, as the paper uses them. pass@k is the chance an agent succeeds at least once in k attempts. pass^k is the chance it succeeds on all k of them. One tells you whether the agent can do the task. The other tells you whether you can rely on it.
The paper's illustration runs each of four tasks five times. Task A succeeds 5 of 5, Task B 3 of 5, Task C 1 of 5, and Task D never. A single attempt lands 45% of the time. Allow five attempts and count any success, and the agent "solves" 75% of the tasks. Either figure could headline a launch chart.
Now picture handing that agent a support queue or a nightly data pipeline. Nobody retries a refund five times and keeps the best one. You need it right on every run, which is exactly what pass^5 counts on those same four tasks.
The authors add a caution I agree with: neither metric shows whether the agent holds up once the environment around it changes.
My take: when a vendor shows you pass@k, ask for pass^k on the same runs. A wide gap between them means you are looking at capability, not reliability, and production only pays for the second.
Quick check before you scroll: Using the four tasks in the paper's example, what does pass^5 come to, and why is it so far below the 75% pass@5 figure?
Full breakdown + the answer: frankduah.me/learnings/2026-10-26-last-week-in-ai-evaluation-and-agentic-ai
If your network builds with AI, a repost puts this in front of people who don't follow me yet.
New here? I post a bite-size AI / ML concept like this every day. Follow me for the daily drop, and it compounds fast. Why I do it: https://lnkd.in/gK8knHDH
#AI #LLM #AIAgents #MachineLearning
The answer
pass^5 comes to 25%. Only Task A succeeded on all five runs, while pass@5 also credits Task B and Task C for landing at least once, so one task out of four counts as reliable against three out of four counted as solved.