Ash Prabaker is a member of technical staff at Anthropic working on reinforcement learning and applied AI, with a focus on autonomous software agents that can complete complex work reliably. Previously a startup founder, he applies experience building AI products to the question of how models can plan, verify and improve software over extended runs.
From finance to AI product development
Prabaker studied economics at the University of Oxford and began his career in financial research at Morgan Stanley. He subsequently worked on coding agents and code search at bloop before co-founding Otio, an AI research assistant. As chief product officer, he helped the startup reach approximately 200,000 users and raise $1.5 million, then joined the founding team at iterate.com to work on coding agents and software for building companies.
At Anthropic, Prabaker has worked with Claude Code, the Claude Agent SDK and startups adopting applied AI. He has also contributed to the behaviors and skills of Claude Mythos Preview and helped advance product work around Anthropic’s partnership with the Atlassian Williams Formula 1 team.
How he thinks autonomous agents should work
In a workshop shared with Andrew Wilson, Prabaker outlined several specific principles for building agents that remain productive over hours:
Adversarial agent evaluation. Coding agents make unreliable judges of their own output. Prabaker separates the builder from an independent critic that opens the running application, tests it through browser automation and returns concrete failures for repair.
Planner–generator–evaluator architecture. A planner establishes broad product direction; the generator and evaluator negotiate testable acceptance criteria before implementation. In one retro-game project, this structure produced working controls and a usable editor where a simpler agent generated an interface that looked finished but could not actually play the game.
Design taste as an evaluation problem. Prabaker grades subjective quality against explicit rubrics covering design, originality, craft and functionality. A sufficiently demanding evaluator can reject generic interfaces and push the builder to abandon weak concepts instead of endlessly refining them.
Persistent artifacts and structured handoffs. Shared files, written acceptance criteria, concise documentation and records of attempted fixes preserve actionable project history across agents and give human developers a practical starting point for subsequent work.
Trace-driven harness design. Prabaker improves agent systems by inspecting where their decisions diverge from human expectations, then adjusting prompts and evaluators. As models improve, he favors removing scaffolding that no longer helps, including unnecessary context resets, rigid task decomposition and excessive evaluation loops.
His public project openapi-to-mcp converts OpenAPI specifications into Model Context Protocol servers, making existing APIs usable through Claude Desktop. He has also taught proactive Claude Code workflows that inspect repositories and initiate useful tasks on a schedule.
In an essay assessing AI’s trajectory, Prabaker extends these technical concerns to falling model costs, changing software development, labor disruption, misinformation and scientific progress. His central challenge is making increasingly autonomous systems capable enough to perform meaningful work and accountable enough to earn trust.
