An AI agent gets one pass-or-fail grade per task — how do you tell which of its many steps were actually smart?

This explores how to score the quality of an agent's individual choices from its recorded runs, without paying humans to label each step. The corpus points to three routes: read the shape of the trajectory, isolate the moments where a choice mattered, and harvest what a deployed system already logs.

The first route uses the trajectory's own structure. An agent usually gets one verdict at the end (it worked or it didn't), which says nothing about which of its fifty steps were smart. Three methods spread that single verdict across the steps by using different structural features. Tree-GRPO uses the branching topology of the run, Supervised RL uses how closely actions match an expert's, and ToolPO uses where tool calls sit. Each turns a sparse end-of-run reward into a step-by-step signal, and none needs a separately trained, hand-annotated process reward model (Can trajectory structure replace hand-annotated process rewards?).

The second route is to find the forks. Most long-horizon benchmarks collapse a run into pass/fail, which hides whether the agent made good decisions or got lucky. Taste-Bench instead isolates decisions at the points where a trajectory branches. The best frontier model scores only 59.7% accuracy on those choices, and the summary says this is done without human annotation (Do long-horizon benchmarks actually measure decision quality?). The broader evaluation literature backs the same move. Two agents with identical success rates can differ enormously in efficiency, reliability, and verification cost (How should we measure agent system performance beyond task success?). Evidence therefore has to come from the whole interaction sequence, scored on process quality and recoverability, not only correctness (How should we evaluate agent behavior beyond final answers?).

The third route is to treat a deployed system as a labeling machine. A routing harness already records execution trajectories, estimates of how demanding each task was, and outcomes. Those can be converted into labeled training examples for fine-tuning and distillation, so the harness doubles as a difficulty labeler (Can a routing harness generate its own training data automatically?). Some trajectory features are also cheap to read off directly. Across 17 frontier models on long optimization tasks, the strongest predictor of success was persistence: repeated benchmark-edit-incorporate cycles, versus quitting early or burning budget without progress (What predicts success in ultra-long-horizon agent tasks?). That is a rough proxy for decision quality that needs no annotator.

The reason to extract this from an agent's own runs is that the alternative has a ceiling. Agents trained only on curated expert demonstrations never learn from their own failures, so their competence is capped by what the curators imagined (Can agents learn beyond what their training data shows?). Trajectory-derived signals let an agent be graded on choices no demonstrator ever showed it. The corpus offers no single solved recipe, though. The methods above give either relative signals (which branch or step did better) or proxies (persistence, expert-alignment), and one of them still leans on expert actions as its anchor.