Prashant Mital

Prashant Mital is a startup solutions architect at OpenAI who helps companies build reliable AI products. His experience spans distributed systems, data infrastructure, enterprise software deployment, and the practical architecture of production agents.

Earlier in his career, Mital contributed to Motor, MongoDB’s asynchronous Python driver, and announced its 2.2.0 release in 2020. At Retool, he became an early deployed engineer, worked with enterprise customers, and established its professional services function.

At OpenAI, Mital advises startups adopting frontier models. With Hillary Bush, he coauthored practical guidance for building with GPT-5, covering the Responses API, evaluation baselines, reasoning controls, and production problems including excessive tool calls, latency, and unnecessarily long answers.

Principles for production agents

• Understand primitives before adding abstractions. Inspect model calls, tool behavior, data flow, and failure modes directly; introduce frameworks when they address a demonstrated engineering need.
• Start with a single-purpose agent. Deploy one focused task to a limited audience, then use retrieval errors, hallucinations, latency, and adoption patterns to guide improvements.
• Add context-preserving agent handoffs selectively. Specialize when workflows genuinely require it: an inexpensive model can triage requests while a stronger reasoning model determines refund eligibility.
• Run guardrails alongside the main workflow. Separate checks can catch prompt injection and unsafe outputs while delaying sensitive actions until safeguards complete. Mital developed these positions during his AI Engineer Summit presentation with Toki Sherbakov.

He also advocates metaprompting: asking a model to draft task-specific instructions, then refining them around the user’s actual priorities, an approach he illustrated through the challenge of comparing complicated leases.