🤖 Agentic coding as pair programming

November 10, 2025

As a product manager with software development history...

I've absolutely vibe coded things...

Even bad things like (blog structure)

But I'm experiencing something new — it's less about vibe and more about context engineering: structuring requirements and constraints so agents produce reliable output.

AND knowing what NOT to do (guardrails)

Example: Asked agent to...

Left, made breakfast, came back reviewed the pull request.

As a product manager, the reduction in time to wait for seeing your gaps in product requirements is incredibly valuable

Before AI I'd write a spec, wait a week for the dev to work on it, maybe ask questions if they weren't introverted or jaded (it wasn't in the spec!), and then see it for the first time during sprint review.

Now I write a spec, ask an agent to build it, and see it within 10 minutes.

If I don't like it, I'll ask it to make updates.

In the project Management triangle there's cost, time and scope - cost and time has gone down dramatically all across the board.

Write documentation with ChatGPT or Goose. Have an autonomous agent like GitHub Copilot or Codex write the code from documentation. Review and refine in the IDE using Claude — a multi-agent orchestration loop where each agent plays a specialized role.

The human-in-the-loop pattern: you're the reviewer, the architect, the product manager — the agents are the pair.

Like twitch for programming

Categories: blog

Tags: technology, systems