Gaurav Mishra outlines how reinforcement learning rewards an agent’s completed trajectory and why verifiable coding tasks are a strong fit. He then argues that coding ability alone is insufficient for computer use. Browser workflows expose agents to incomplete observations, expired credentials, unpredictable loading behavior and actions that can have lasting consequences.
Gaurav Mishra presents early training trajectories in which an expense-submission agent guesses passwords until an account is blocked and follows an advertisement’s misleading submit button. He uses these examples to distinguish reaching a nominal outcome from completing the intended task safely. His training approach includes realistic interface disruption, recovery actions, process-level penalties for dangerous behavior and calibration of when to ask a user for help.
Gaurav Mishra separates improvements to the model from protections in its surrounding harness. The model needs screen grounding, change detection and an understanding of the limits of screenshots and page structure. The harness can add risk classification, credential checks, monitoring for unproductive loops, audit trails, checkpoints where recovery is possible and enforced human handoff.
Gaurav Mishra closes with a later training trajectory that distinguishes the real submit button, recognizes an expired session and hands off rather than guessing credentials. A user-simulator agent supplies the login before the original agent resumes the expense task. He describes a development loop that gathers failures from controlled use, strengthens training and relies on runtime safeguards while model capabilities improve; the example is not evidence that all deployment risks are solved.
Watch on YouTube
