The paper proposes a confidence‑calibration technique for LLM‑based agents that aligns the model’s self‑reported verbalized confidence with the eventual success or failure of a multi‑step episode rather than with the correctness of each individual action. Using GRPO, the set of rollouts generated for a given task provides binary outcome labels, and a proper scoring rule evaluates each step’s confidence against that episode‑level result. In offline reinforcement learning on trajectories collected from a Qwen2.5‑7B agent operating in ALFWorld, TextWorld, and WebShop, the area under the ROC curve for step‑wise confidence versus episode outcome improves from chance levels to between 0.68 and 0.85, and the decile of steps the agent rates most confidently succeeds roughly 4.3 times more frequently than the average step. Online RL on a live ALFWorld agent employing four calibration rewards further raises early‑step AUROC, reduces expected calibration error from 0.42 to under 0.12, and yields better selective execution than both the untrained baseline and a format‑only RL control while preserving overall task success. An accompanying evaluation protocol that disentangles episode depth, task difficulty, and attempt‑level foresight shows that the calibrated agent can often predict failure from its first step, whereas distinguishing successful from unsuccessful attempts on the same task remains challenging, with only modest early‑window signals appearing in TextWorld.