Gaurav Mishra is a researcher on Amazon’s AGI Autonomy team, developing computer-use agents that can navigate real software while recognizing the limits of their authorization. His central challenge is translating reinforcement-learning advances into reliable autonomy amid expired credentials, deceptive interfaces, irreversible actions, and incomplete information.
Mishra spent a decade at Google, initially working on Google Search indexing and infrastructure for deploying language models. He joined Google Brain in 2019 and spent approximately six years across Brain and Google DeepMind, contributing to large-scale pretraining, early language models, Google Gemini, and Gemini training for Search AI Mode. His account of his move to Amazon outlines that progression from search infrastructure to frontier models and autonomous agents.
At Amazon, he focuses on systems that extend coding-agent capabilities into browsers and other applications. Reinforcement learning works particularly well when tasks have verifiable outcomes and multiple valid solutions, but computer use introduces problems that coding benchmarks obscure: screenshots and browser markup provide incomplete information, interfaces change unexpectedly, and completing an assignment can still produce unauthorized consequences.
Making autonomy safe enough to deploy
Mishra’s approach combines improvements to training, perception, and execution:
- Realistic digital sandboxes: Simulated environments should include slow pages, misleading advertisements, pop-ups, missing labels, interrupted sessions, and shifting layouts.
- Recovery-aware computer-use agents: Agents should learn to refresh, backtrack, wait, abandon unsafe paths, or escalate instead of assuming failures can simply be reset.
- Process reward models: Training should assess the entire action sequence, penalizing risky detours and side effects even when an agent completes its assigned task.
- Multisource visual grounding: Agents must integrate screenshots, browser structure, interface semantics, and visible changes to identify legitimate controls and recognize deceptive alternatives.
- Human handoff for consequential actions: Execution safeguards should classify action risk, check credentials, maintain audit logs, and return control when authorization expires or intervention is necessary.
In a browser-agent demonstration, Mishra contrasted an early system that guessed passwords or followed a sponsored lookalike button with an improved agent that recognized the deception, requested human authentication, and resumed filing an expense. His deployment strategy starts with strong execution safeguards, then incorporates observed failures into training so protective scaffolding can narrow as the underlying model becomes more reliable.
