--- title: From RL to IRL — Gaurav Mishra, Amazon AGI Lab | daily.dev description: A researcher at Amazon's AGI Lab explains why reinforcement-learning-trained coding agents struggle when deployed as computer-use agents in the real world. RL... canonical: https://daily.dev/posts/from-rl-to-irl-gaurav-mishra-amazon-agi-lab-f4xgvpo8o twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: From RL to IRL — Gaurav Mishra, Amazon AGI Lab | daily.dev og:description: A researcher at Amazon's AGI Lab explains why reinforcement-learning-trained coding agents struggle when deployed as computer-use agents in the real world. RL... og:url: https://daily.dev/posts/from-rl-to-irl-gaurav-mishra-amazon-agi-lab-f4xgvpo8o og:image: https://api.daily.dev/og/posts/f4xgvPO8o.png og:image:alt: From RL to IRL — Gaurav Mishra, Amazon AGI Lab og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # From RL to IRL — Gaurav Mishra, Amazon AGI Lab **[AI Engineer](https://daily.dev/sources/aidotengineer)** · 17 min read · 0 upvotes · 0 comments ## Summary A researcher at Amazon's AGI Lab explains why reinforcement-learning-trained coding agents struggle when deployed as computer-use agents in the real world. RL works well for coding because tasks are verifiable, but real-world computer use introduces partial observability, irreversible actions, non-determinism, expiring credentials, ambiguous success criteria, and adversarial UI content (like fake buttons). Demos show early agents guessing passwords, locking accounts, or clicking sponsored buttons that lead to leaking personal data. The proposed fix is a 'flight school' training approach: high-fidelity messy simulations, process reward models that penalize risky steps not just outcomes, calibrated confidence for escalation to humans, adversarial training tasks, upgraded perception in the model, and a harness layer providing checkpointing, rollback, credential guardrails, execution monitoring, audit logs, and human handoff. ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Questions this post answers ### why do reinforcement learning trained coding agents fail when used as computer use agents in the real world They fail because real environments break RL's core assumptions: state is only partially observable through screenshots and DOM, actions like submitting forms or deleting files are often irreversible, credentials expire mid-task, success is ambiguous, and pages contain adversarial content like fake buttons designed to mislead. In demos, agents guessed at expired passwords until accounts got blocked, or clicked a sponsored button resembling the real submit button and began leaking personal details on the wrong site. _daily.dev surfaces engineering writeups like this for teams debugging why their own agents misbehave in production._ ### how can you train an ai agent to handle expired login credentials without breaking the task Instead of guessing a new password or resetting state, the agent should recognize the expired session, halt the task, and hand off control to a human or user-simulator to re-authenticate, then resume with prior context (like a preserved form amount) intact. This requires calibrated confidence in the harness so the model knows when to escalate rather than attempt risky recovery actions on its own. _engineers building resilient agent harnesses track these recovery patterns on daily.dev._ ### what components make up a training harness for computer use ai agents A harness handling computer-use agents typically includes checkpointing and rollback for risky states, an action risk classifier to flag unsafe proposed actions, credential guardrails that detect expired sessions, an execution monitor that catches loops or unproductive repeated clicks, audit logs of all actions and effects, and a human handoff mechanism when the model's confidence calibration is insufficient. _teams designing agent guardrails follow deep dives like this on daily.dev._ ## Similar posts on daily.dev - [How do AI teams use RL in production?](https://daily.dev/posts/how-do-ai-teams-use-rl-in-production--84y0ggetl) · Daily Dose of Data Science \| Avi Chawla \| Substack · 2 upvotes · 1 comments - [Why AI agents are like the dog that pushed kids into the Seine](https://daily.dev/posts/why-ai-agents-are-like-the-dog-that-pushed-kids-into-the-seine-17cvyhnc2) · CSO Online · 0 upvotes · 0 comments --- Tags: [#llm](https://daily.dev/tags/llm), [#ai-agents](https://daily.dev/tags/ai-agents), [#automation](https://daily.dev/tags/automation), [#amazon](https://daily.dev/tags/amazon), [#reinforcement-learning](https://daily.dev/tags/reinforcement-learning) [View this post on daily.dev](https://daily.dev/posts/from-rl-to-irl-gaurav-mishra-amazon-agi-lab-f4xgvpo8o) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"From RL to IRL — Gaurav Mishra, Amazon AGI Lab","url":"https://daily.dev/posts/from-rl-to-irl-gaurav-mishra-amazon-agi-lab-f4xgvpo8o","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/from-rl-to-irl-gaurav-mishra-amazon-agi-lab-f4xgvpo8o"},"datePublished":"2026-08-14T16:03:11.852Z","dateModified":"2026-08-18T15:18:24.832Z","description":"A researcher at Amazon's AGI Lab explains why reinforcement-learning-trained coding agents struggle when deployed as computer-use agents in the real world. RL...","image":"https://i.ytimg.com/vi/Cc0_nyxROBA/sddefault.jpg","thumbnailUrl":"https://i.ytimg.com/vi/Cc0_nyxROBA/sddefault.jpg","isAccessibleForFree":true,"articleSection":"AI Engineer","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"AI Engineer","logo":"https://media.daily.dev/image/upload/s--u5PucxNT--/f_auto/v1724338940/logos/aidotengineer","url":"https://daily.dev/sources/aidotengineer"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/from-rl-to-irl-gaurav-mishra-amazon-agi-lab-f4xgvpo8o","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":0},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"llm,ai-agents,automation,amazon,reinforcement-learning","timeRequired":"PT17M","video":{"@type":"VideoObject","name":"From RL to IRL — Gaurav Mishra, Amazon AGI Lab","description":"A researcher at Amazon's AGI Lab explains why reinforcement-learning-trained coding agents struggle when deployed as computer-use agents in the real world. RL...","thumbnailUrl":"https://i.ytimg.com/vi/Cc0_nyxROBA/sddefault.jpg","uploadDate":"2026-08-14T16:03:11.852Z","duration":"PT17M","url":"https://api.daily.dev/r/f4xgvPO8o","embedUrl":"https://www.youtube.com/embed/Cc0_nyxROBA"}} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"AI Engineer","item":"https://daily.dev/sources/aidotengineer"},{"@type":"ListItem","position":3,"name":"From RL to IRL — Gaurav Mishra, Amazon AGI Lab"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/from-rl-to-irl-gaurav-mishra-amazon-agi-lab-f4xgvpo8o#faq","mainEntity":[{"@type":"Question","name":"why do reinforcement learning trained coding agents fail when used as computer use agents in the real world","acceptedAnswer":{"@type":"Answer","text":"They fail because real environments break RL's core assumptions: state is only partially observable through screenshots and DOM, actions like submitting forms or deleting files are often irreversible, credentials expire mid-task, success is ambiguous, and pages contain adversarial content like fake buttons designed to mislead. In demos, agents guessed at expired passwords until accounts got blocked, or clicked a sponsored button resembling the real submit button and began leaking personal details on the wrong site. daily.dev surfaces engineering writeups like this for teams debugging why their own agents misbehave in production."}},{"@type":"Question","name":"how can you train an ai agent to handle expired login credentials without breaking the task","acceptedAnswer":{"@type":"Answer","text":"Instead of guessing a new password or resetting state, the agent should recognize the expired session, halt the task, and hand off control to a human or user-simulator to re-authenticate, then resume with prior context (like a preserved form amount) intact. This requires calibrated confidence in the harness so the model knows when to escalate rather than attempt risky recovery actions on its own. engineers building resilient agent harnesses track these recovery patterns on daily.dev."}},{"@type":"Question","name":"what components make up a training harness for computer use ai agents","acceptedAnswer":{"@type":"Answer","text":"A harness handling computer-use agents typically includes checkpointing and rollback for risky states, an action risk classifier to flag unsafe proposed actions, credential guardrails that detect expired sessions, an execution monitor that catches loops or unproductive repeated clicks, audit logs of all actions and effects, and a human handoff mechanism when the model's confidence calibration is insufficient. teams designing agent guardrails follow deep dives like this on daily.dev."}}]} ```