Aditya Bhargava is a staff software engineer at Etsy, the author of Grokking Algorithms, and the creator of Agency, a programming language for building AI agents. His work combines illustrated technical education with a distinctive approach to agent design: better software infrastructure can make autonomous systems safer, more capable, and less dependent on powerful proprietary models.
Bhargava’s background spans computer science and fine arts. In 2013, he published an illustrated guide to functors, applicatives, and monads, applying visual explanation to notoriously difficult functional-programming concepts. His book Grokking Algorithms, first published in 2016 and updated in 2024, brought that approach to searching, sorting, graphs, and computational tradeoffs. He later extended it to cloud networking, OAuth, and linear algebra through DuckTyped.
At Etsy, Bhargava has worked on localization and translation, moved into engineering management, and subsequently held a staff engineering role. He has also served as an individual-contributor lead for agentic commerce, developing systems that act on users’ behalf.
What makes his agent work distinctive
- The agent harness shapes model performance. Bhargava treats tools, permissions, execution controls, and evaluation as consequential engineering problems. His agent-harness research argues that stronger infrastructure could help smaller, locally runnable models handle tasks otherwise reserved for expensive proprietary systems.
- Agency makes agent safety programmable. His open-source Agency language compiles to TypeScript, converts ordinary functions into model-callable tools, and generates their schemas automatically. Interrupts and resumable execution enable human approval before sensitive operations.
- Partial function application creates bounded autonomy. Fixing a tool’s permitted directory before an agent receives it restricts filesystem access without requiring approval for every operation.
- Feedback makes agents measurable. A ReAct-style feedback loop lets coding agents inspect failures, modify programs, and verify repairs against tests. Specialized subagents separate unrelated capabilities, while built-in optimization improves prompts against explicit objectives.
