Part 1 of 6 · 1 min read
Overview
Agents face choices at every level: attempt a task again, work on this ticket before that one, route to prompt A or prompt B. Each is cheap to describe and expensive to get wrong, because it multiplies model calls or hides a bias. This is a Specialist-tier module: the aim is to give you the vocabulary and the cost arithmetic, so you can answer "how would you use extra compute?" and "how does an agent learn which strategy works?".
About the example, and what is real
Everything runs offline in code/mini-explore. Real: the sampling, knapsack and bandit code, an exact 0/1 knapsack as the benchmark for prioritisation, the standard bandit algorithms (epsilon-greedy, UCB1, Thompson sampling with Beta posteriors), seeded Monte Carlo, and the 21 tests. Assumed: each task's chance of a correct sample is drawn from a Beta distribution with mean 0.55 (so some tasks are nearly unsolvable); half of wrong answers go to one "attractor" answer; the verifier approves 95 percent of correct and 10 percent of wrong answers; task values are uniform from 1 to 10 and costs integers from 1 to 8; five strategies succeed 30, 35, 50, 55 and 62 percent of the time. No model is called. These show the shape of each trade-off, not typical agent performance.