A situation I have seen more than once lately: two developers are pairing, one has the keyboard and is busy prompting an AI assistant, reading its suggestions and accepting them. The other one nods along. From the outside it looks like pair programming. In reality, one person is working with their AI and a second person is watching.
That made me ask a simple question: if an assistant already reads along and suggests code, who should actually sit at the pairing table? I explore it in detail in Pair programming with AI: roles for humans and assistants. Here are the main ideas.
What an assistant can and cannot take over
Birgitta Böckeler of Thoughtworks broke down the benefits of pairing on martinfowler.com in 2023 and checked how many of them a coding assistant covers. Her answer: only a small part, and I agree. An assistant is great at explaining unfamiliar APIs, getting you unstuck and producing routine code. What it cannot do is carry tacit knowledge: why a field has a strange name, which approach already failed last year, which department gets nervous about which change. That knowledge spreads from head to head as a side effect of pairing.
There is also shared responsibility. When one person writes code alone with an assistant, exactly one human can explain it afterwards. That is a step backwards for your bus factor. So my conclusion is not "assistant or pairing", but the assistant as a third seat inside the pair.
A clear split of roles
Once an assistant joins, the pair needs a new agreement. The split that works for me:
- Driver: types and writes the prompts, and reads every suggestion out loud before accepting it.
- Navigator: holds the goal, checks suggestions against architecture and team agreements, and has the final say together with the driver.
- Assistant: delivers drafts, tests and explanations, but never decides anything.
Two small rules make a big difference. First, when you switch roles, the right to talk to the assistant switches too, so nobody becomes the permanent prompter. Second, for juniors in strong-style pairing, an AI suggestion is only accepted once the driver can explain it in their own words.
A short, verifiable instruction keeps the assistant in its lane, for example after the navigator has written a failing test:
Implement extendLoan in src/loans/extend.ts so that the tests in
extend.test.ts pass. Return a new object instead of mutating the
input. No new dependencies. Then explain in two sentences how you
handled the date calculation.
Small steps, a commit after every green test, and a quick check together: does it fit the goal, do we understand every line, would we approve this in a review?
Feeling fast is not being fast
One finding I keep coming back to: a randomized study by METR from 2025 with 16 experienced open-source developers found they took 19 percent longer with AI tools, while estimating afterwards that they had been about 20 percent faster. The study has clear limits, and METR itself warns against generalizing. Still, it is a good reason not to judge your pairing experiments by gut feeling alone.
For managers, the 2025 DORA report describes AI as an amplifier of what an organization already is. A team that collaborates well will likely benefit. A team where everyone works alone will drift further apart. That is why I would not mandate anything. I would suggest a time-boxed experiment and look at lead time, defects and how many people can explain a given piece of code, not at accepted suggestions.
What the full article covers
- Three setups (solo, pair, mob with an assistant) and their typical risks
- A detailed breakdown of which pairing benefits an assistant covers and which stay human
- A complete two-hour session plan, including an
AGENTS.mdexcerpt and a ping-pong test example - When to close the assistant: architecture decisions, domain onboarding, uneven pace
- Common mistakes such as prompt ping-pong, plus an FAQ on juniors and Copilot
Read it here: Pair programming with AI: roles for humans and assistants
📖 This is the condensed version. The full article on sascha-ms.de goes deeper – with all code samples, diagrams and a short knowledge check.
👉 Next step: More for team leads: my cheat sheets put the essentials on one printable page.
📬 Stay updated: New articles on serverless, AI and leading tech teams via RSS, browser push or the weekly newsletter (in German).
Originally published on sascha-ms.de. More articles on serverless, AI in everyday development and leading tech teams: sascha-ms.de/en
Top comments (0)