Last updated: June 17, 2026
Image: Peppe Silletti / LinkedIn
Senior developers can now offload boilerplate and first-pass review to an AI agent, and ship measurably faster. Junior developers using the same tools without human guidance risk applying confident-sounding suggestions to problems they do not fully understand yet. That practical divide is what the AI pair programming replacement debate keeps circling without naming directly.
The conversation has moved from theoretical to operational. Most developers are not particularly upset about it. Understanding the actual trade-off, though, matters for how you structure your team.
What Pair Programming Actually Got Right
Image: Peppe Silletti / LinkedIn
Pair programming – two developers sharing a single workstation, one typing (“driver”) while the other reviews (“navigator”) – sounds inefficient on paper. Two salaries, one output stream. Management has always hated the maths.
But Peppe Silletti, writing on LinkedIn, is honest about this: pair programming genuinely worked. It caught errors in real-time, provided continuous quality checks that did not require waiting days for a pull request review, and transferred institutional knowledge by osmosis rather than by document. The navigator’s fresh eyes spotted the bug the driver was too close to see.
Yaniv Preiss goes further. After switching his teams to full pairing, he reported his organisation’s DORA metrics – a standardised framework measuring software delivery performance across deployment frequency, lead time, failure rate, and recovery time – nearly doubled. Engineers grew across mobile, frontend, backend, and infrastructure simultaneously. That is a striking claim from a practitioner, though it comes from a single team’s experience rather than a controlled study. The direction is still consistent with what pairing advocates have argued for years: the benefits were not confined to code quality. They extended across the entire delivery pipeline.
The Case for Coding Agents
Here is what changed. AI coding agents – tools like Copilot, Cursor, and Claude’s agentic mode – have made the navigator role available on demand, at zero marginal cost, with no scheduling overhead. Research (not yet peer-reviewed) suggests AI coding assistants reduce task completion time by 20-56% on routine coding tasks, with the strongest gains on boilerplate-heavy work. The agent does not context-switch from its own work. It does not go on holiday. It does not have opinions about your variable naming conventions. (Well. It does. But you can tell it to stop.)
Take a concrete example. A senior engineer building a standard REST endpoint – user creation, validation, database write, response schema – defines the shape in a comment and lets the agent scaffold the controller, routes, validation layer, and test stubs. Review, adjust, ship. What used to take a morning now takes under an hour. The agent does not replace judgment; it eliminates the typing.
Revisiting Using AI Coding Assistants: You’re Holding It Wrong Edition covers why most developers underuse these tools – the agent workflow requires you to direct it like a senior, which is a learned skill.
But Silletti’s deeper argument is not really about coding speed. Silos and handoffs – not slow typing – kill development velocity. A feature dies a dozen small deaths waiting for design sign-off, a product ticket, a backend engineer to unblock a frontend engineer. An AI agent generating code faster does not fix any of that. As explored in agentic development hinges on verification for cloud-native teams, the harder constraint is knowing what to verify – which requires judgment, not speed.
What Silletti actually proposes is more radical: apply the pair programming philosophy to the entire product lifecycle. Pair on user interviews. Pair on prioritisation sessions. Pair on sketches. Build teams of cross-functional “builders” who own problems end-to-end rather than hand them off across functional boundaries.
Where AI Agents Fall Short of Human Pairing
You might think the evidence here is clear-cut. It is not. A 2026 systematic review of 48 peer-reviewed publications on AI-native software engineering found the evidence base internally contradictory – studies disagree on both the magnitude and the direction of productivity effects from AI coding tools. Some show significant gains. Others show near-zero impact or increased error rates. Task type, developer experience level, and the specific tool all shift the results substantially.
Ed Hodapp puts the sharpest point on the real trade-off: pair programming still matters for instilling systems knowledge in junior engineers. Using an AI agent effectively is like being a tech lead – you need to know what questions to ask, when to trust the output, and when the confident-sounding suggestion is wrong.
Here is where that gap becomes costly. A junior engineer is asked to refactor a data-fetching layer. The agent suggests extracting a custom hook – clean, idiomatic-looking React. The junior applies it without questioning. The hook creates a new subscription on every render because the dependency array is wrong. The agent did not know the team’s state management conventions or the performance constraints upstream. The junior did not have the systems knowledge to spot the error. A senior navigator would have caught it in thirty seconds by asking “what does the parent component do with this data?”
That is the failure mode of AI pairing for juniors: the agent generates plausible code, not necessarily correct code for your specific context. Research on AI coding assistants notes the strongest gains for junior developers on routine tasks – but “faster” is not the same as “right” when the task involves judgment about systems you have not yet internalised.
AI-first teams are restructuring how they onboard engineers, but the fundamentals of knowledge transfer have not changed: juniors learn best alongside people who can articulate their thinking in context, not just generate the answer.
When to Pick Each Approach
If your team is experienced and the main drag is handoffs rather than code review lag, AI agents are the right investment. Break down the silos Silletti identifies – merge product, design, and engineering responsibilities where you can, and let agents handle the navigator role for code. The pair programming ritual becomes unnecessary overhead.
If your team includes junior engineers who need to internalise systems knowledge, use AI agents as a supplement, not a replacement for human pairing. The agent accelerates execution; the experienced human navigator builds mental models and – crucially – explains the reasoning behind decisions, not just the decisions themselves. One without the other leaves gaps that surface later, usually at the worst moment.
Pair programming did not die because developers stopped valuing collaboration. It died because most organisations would not pay for it. AI agents have made one specific implementation obsolete. The underlying question is the same one it always was: where is your team actually losing time? Handoffs and silos need structural change. Code review lag on a senior team? Your AI navigator is already waiting.
Frequently Asked Questions
Q: Has AI fully replaced pair programming?
A: AI coding agents have made the traditional navigator role available on demand at no marginal cost, effectively replacing pair programming for many experienced teams. However, for junior developer onboarding and systems knowledge transfer, human pairing still offers advantages AI tools cannot replicate.
Q: What are DORA metrics and why do they matter here?
A: DORA metrics measure software delivery performance across deployment frequency, lead time for changes, change failure rate, and time to restore service. One practitioner reported nearly doubling these metrics after adopting full pairing – a strong signal, though from a single team’s experience rather than a controlled study.
Q: Is the evidence for AI coding tools improving productivity conclusive?
A: No. A 2026 systematic review of 48 peer-reviewed publications found the evidence base internally contradictory – studies disagree on both the size and direction of productivity effects. Separate research suggests routine task completion time drops by 20-56%, but results vary significantly by task type, experience level, and tool.
Q: Should junior developers use AI agents instead of pairing with senior engineers?
A: Not as a direct replacement. Directing an AI agent effectively requires the same judgment as being a tech lead – knowing what to ask, when to trust the output, and when it is wrong. Junior developers are still building those mental models, and a human navigator who explains reasoning in context is more effective at that stage.
Q: What does Silletti actually propose as an alternative to traditional pair programming?
A: Silletti argues the pair programming philosophy should be applied to the entire product lifecycle – pairing on user interviews, prioritisation, and design sketches, not just on code. He advocates for cross-functional “builder” roles that collapse the silos between product, design, and engineering, since those handoffs are the real source of lost delivery time.
This article was researched and written with AI assistance, then reviewed for accuracy and quality. Nia Campbell uses AI tools to help produce content faster while maintaining editorial standards.
Need help with your web project?
From one-day launches to full-scale builds, DRS Web Development delivers modern, fast websites.
