Jared Zoneraich is a builder in residence at Cognition, where he works on the AI software engineer Devin, and the co-founder and former chief executive of PromptLayer. He has built products around two closely related challenges: helping teams control how language models behave and giving autonomous coding agents the tools to perform useful software work.
From developer tools to coding agents
Zoneraich founded the high-school hackathon hackBCA and studied electrical engineering and computer science at the University of California, Berkeley. He subsequently worked at Google, Facebook, and machine-learning infrastructure company Tecton, and was an associate at venture firm Alsop Louie Partners.
With Jonathan Pedoeem, he co-founded PromptLayer, developing a prompt registry and version-control system alongside tools for evaluation, request logging, experimentation, and production monitoring. The company announced a $4.8 million seed round in 2025. Zoneraich also maintains the PromptLayer Python package.
His product philosophy centers on collaboration between engineers and domain experts: lawyers, physicians, product managers, and other specialists should help shape prompts and evaluate outputs in fields where they understand quality better than the engineers integrating model APIs.
Simple agent loops, rigorous tools. Zoneraich favors a capable model repeatedly selecting tools and responding to their results over elaborate orchestration graphs. Exploration can remain flexible while sensitive operations and predictable deliverables are handled by tightly specified, independently tested tools.
Bash as a universal agent interface. His case for virtual-machine-backed agents emphasizes shell access, isolated execution, and persistent filesystems. Familiar command-line workflows let models inspect repositories, run tests, write temporary scripts, and recover from errors without requiring dozens of specialized integrations.
Context management through isolated sub-agents. Dedicated research, documentation, testing, and review agents can operate in separate contexts and return only relevant findings. Files, summaries, and selectively loaded skills keep the main agent focused during extended work.
Agent evaluation grounded in real work. Zoneraich prioritizes historical backtesting, end-to-end trials, targeted conversation snapshots, tool-call patterns, and LLM-as-judge evaluation over generic benchmark claims. His analysis of coding-agent architecture compares Claude Code, Codex, and other independently developed products without claiming authorship of them. One practical workflow has an agent inspect recent repository changes, update documentation, and submit a pull request for human review.
