Mohit Aggarwal: What autonomous AI agents mean for product managers

Mohit Aggarwal, a product manager and writer on AI, published this article on Medium in April 2026. The piece is prompted by a specific observation: watching an AI agent open a spreadsheet, pull data from three separate sources, build a pivot table, format the result, and post it to Slack — without a single prompt between each step. No handholding, no confirmation requests, just sequential execution of a multi-step task that previously required a person.

The article locates this shift at the intersection of two infrastructure developments. MCP (Model Context Protocol) reached 97 million installations, providing the connective layer that lets AI agents talk to tools like databases, project management software, and communication platforms. Separately, GPT-5.4 gained the ability to directly operate a computer, not through API calls but by navigating a desktop the same way a human would. Together, these two changes moved the capability boundary from “AI can help me with a task” to “AI can execute a task I would otherwise need to delegate to a person.”

The article does not treat this as a technical curiosity. Aggarwal argues that the standard PM toolkit — sprint planning, user flow design, feature specifications — was built around the assumption that a human would execute whatever was decided. When an agent can handle the execution of complex workflows autonomously, the design decisions embedded in those workflows need to be reconsidered. Who is accountable when an agent routes a support ticket incorrectly? How do you spec a feature whose core logic runs inside an agent that is also allowed to modify its own behavior?

The piece is not prescriptive about what product managers should do in response. It reads more as a diagnostic — an attempt to document a threshold that Aggarwal thinks most PMs have not yet recognized. That framing makes it useful for PMs at companies actively building agentic products or integrating autonomous agents into internal workflows, where the questions it raises are already becoming concrete.