The AI-Native Individual Contributor

The individual contributor’s job has been, for as long as the title has existed, mostly production. You were hired to write the code, draft the memo, build the model, work the queue, and seniority was the slow accumulation of having done enough of it to know what good looks like. That arrangement is coming apart. The production is being automated, and what is left is harder to hire for, harder to teach, and worth more: deciding what gets made, directing the machines that make it, and being the one accountable when it is wrong.

I run an operation where the default actor for any incoming piece of work is an agent. The humans are few and each is high-leverage. They set intent, route the work, judge what comes back, and own the outcome. I want to be precise about what that role actually is, because the popular version — “you’ll just manage AI” — hides everything that matters. The scarce skill is the encoded logic that decides which agent handles what and when a human has to step in. That logic is the durable asset, and the models underneath are rented and swappable.

The gains are real, and they move the cost from writing to checking

The optimistic case is genuine, and the richest of the field experiments states it best. A study run with Boston Consulting Group gave 758 management consultants realistic tasks chosen to sit inside AI’s competence, and the consultants using GPT-4 finished 25.1% faster while producing work rated more than 40% higher in quality than the control group’s.² The other trials confirm the direction. A randomized study of GitHub Copilot had developers build an HTTP server from scratch and found the tool’s group finishing 55.8% faster,¹ and a study of over five thousand customer-support agents found AI lifting average resolutions per hour, with the gains running heaviest to the least experienced workers.³

Then a separate randomized trial ran the question in a harder setting and got the opposite sign. METR took experienced open-source developers working in mature codebases they knew well, with an average of five years on those exact projects, and measured them on real tasks. With AI tools allowed, they were 19% slower than without them.⁴ The study also asked everyone to predict the effect. Before starting, the developers expected the tools to speed them up by 24%, and after finishing — after living the slowdown — they still believed the AI had made them about 20% faster.⁴ Economics and machine-learning experts asked to forecast the result predicted speedups near 40%.⁴ Everyone, including the people doing the work and measuring themselves, was wrong by a wide margin and in the same direction.

Together the two trials say something neither says alone. In the clean greenfield task, writing the code was the work, and the machine did it. In the mature codebase, writing was never the bottleneck; understanding the system, integrating cleanly, and meeting the quality bar were. AI made the writing cheaper and the checking more expensive, and the second cost was larger. The developers could not feel it because the tool felt productive the whole time. It produced. It just produced things that had to be checked, and the checking ate the saving.

AI failing in hard settings is the smaller lesson here. The larger one is that the cost moved. Production gets cheap, verification becomes the binding constraint, and verification is exactly the thing people are worst at judging in themselves.

The gains stop at an invisible edge, and confidence carries people past it

The consulting experiment ran one more task, deliberately chosen to sit just outside the model’s competence: similar in apparent difficulty, past the edge of what it could actually do. On that task the consultants using AI were 19 percentage points less likely to reach the correct answer than those working without it.² The tool did not refuse or stall. It produced a confident, plausible, wrong answer, and the humans took it.

Ethan Mollick, a co-author of that experiment, calls the boundary the jagged frontier. Tasks of near-identical apparent difficulty fall on opposite sides of an invisible line of AI capability.⁵ Inside the line you get the lift. Step outside without noticing, and the same fluent confidence that helped you now degrades your work, because you trust it exactly as much. Learning where the edge runs takes sustained practical use,⁵ and that learned map is the thing the experienced person carries and the tool cannot supply.

A small longitudinal study put numbers on what happens when the edge goes unwatched. Over six months, participants leaned on AI most heavily for the hardest tasks, at 73.9% reliance, while their actual accuracy collapsed as problems got harder — 95.2% correct on the simplest, 47.8% on the most complex.⁶ On those complex problems they overrated their own correctness by 34.6 percentage points. They believed they were right and they were wrong, and the belief was strongest exactly where the performance was weakest.⁶ It is a small academic pilot and the magnitudes should be read as indicative, not as workplace estimates.⁶ But the mechanism is the same one the developers and consultants ran into, measured a third way. Confidence and correctness come apart, and the gap between them widens precisely where it is most dangerous.

Verification has to be built as a separate layer

The standard advice, stay critical and check the AI’s work, is useless here, because the evidence says people cannot tell when they have failed to. You cannot solve a metacognitive blind spot with more diligence. The blind spot is the diligence not firing.

So verification has to stop being a posture and become a part you build. In my operation it is a distinct layer: an independent measure of answer quality, separate from the system that produced the answer, observable and logged, with its own job. It never asks the model how sure it is, because high confidence is not a safety signal. There are two cases the layer has to catch. The AI saying “I’m not sure” is the easy one, since the uncertainty is visible. The AI being confidently wrong is the hard one, because catching it requires a check that does not share the cognition that made the error. And the two turn out to be one skill; competence-detection and error-detection are one capability.

This reframes what an individual contributor is for. The machine produces the work, and you are the layer that catches the plausible-wrong answer before it reaches anyone, and the one accountable when it does anyway. Simon Willison draws the line between vibe coding, where you stop caring what code comes out, and what he calls vibe engineering, the professional accelerating with the machine while staying “proudly and confidently accountable” for what ships.⁷ If a model wrote every line but you reviewed, tested, and understood all of it, then the model was a typing assistant and you still held the pen.⁷

There is a structural trap waiting here. Willison notes that fast production makes review the bottleneck, and bottlenecks are what organizations reflexively remove.⁷ The field data shows the trap already sprung. One industry report found pull-request review times ballooning 91% as AI-written code flooded in, individual speed that never became team speed, because the verification downstream was never resourced to match.⁸ Addy Osmani, synthesizing the productivity evidence, puts the real gains around 20 to 30%, a long way from the promised 10x, and finds the dominant time sink in “almost-correct” output, the plausible-but-flawed suggestion that shifts your effort from writing to debugging.⁸ Obviously-wrong output is free to discard. The nearly-right output is what costs you, and producing more of it faster is a gain only if the checking keeps up.

Microsoft’s 2026 Work Trend Index, a vendor survey of 20,000 AI-using workers and directional rather than causal evidence, finds the same pattern at population scale. The teams getting the most from AI stand out on team practices — they discuss quality standards for AI-assisted work at nearly twice the rate of everyone else — and organizational factors carry roughly twice the AI impact of individual effort alone.¹¹ Verification capacity, on this evidence, is a team property before it is a personal one.

The work moves from producing the answer to routing the question

If the human is no longer producing, the day-to-day work is routing: deciding which agent, or which person, handles a given piece of work, with what brief, and when it has to escalate.

The pattern has a documented engineering form. Anthropic’s multi-agent research system runs on an orchestrator-worker design, in which a lead agent analyzes the request, decomposes it, spawns specialized subagents to work in parallel, and synthesizes what comes back, giving each worker a clear objective, an output format, the right tools, and explicit boundaries.⁹ The lead agent’s value sits in the decomposition and the briefing rather than in doing the subtasks itself. The same team is explicit that output quality depends on a deliberate evaluation step, an LLM judge scoring against a rubric for accuracy, completeness, and sourcing, plus human evaluation for the edge cases the automation misses.⁹ That is one vendor describing its own product, and I would not rest the argument on it alone. What I can add is that it matches the operation I actually run.

Here is the routing decision at its most ordinary. A question arrives that reads as routine, the kind an agent closes in seconds. The routing layer checks it against the domains the agents are cleared to answer and against what the detection layer reports about the question itself, and most of the time it passes, an agent answers, and no human touches it. The case the layer exists for is the one that reads the same and is different — a question whose phrasing sits inside an agent’s competence and whose substance sits outside it. If that route goes through, the agent will answer fluently, and in a regulated domain an answer that should never have been given is a compliance event whether or not anyone catches it later. So the escalation is the layer’s real product. Deciding what an agent may close alone, what goes to a person, and what triggers a second check is the encoded judgment earning its keep, one routed question at a time.

Workers already using AI describe their own posture in these terms. In the Microsoft survey, 86% of AI users say they treat AI output as “a starting point, not a final answer”¹¹ — the verify-and-direct stance, self-reported at scale.

One unglamorous discipline holds the whole thing up, and it is the easiest to skip. You have to measure agent output quality continuously, against domain-specific criteria, at scale. The alternative is catching one bad case when someone happens to notice it, and “we’ll notice” is not a mechanism. A model swap can change output quality invisibly, drift accumulates below the threshold anyone would flag by eye, and the verification research already says people cannot feel the degradation even in their own work. Continuous evaluation is the standing capability that replaces “we’ll notice” with something that actually fires, and it is both the least glamorous part of the role and the part the role cannot do without.

There is a reading of the evidence that cuts against this whole operating model, and it deserves stating plainly. The METR result can be read as an instruction to keep experienced humans producing in mature, high-stakes settings; the tools slowed those developers down, so leave the production with the people who know the system. For some work, today, that is the right call. I still build agent-default, and my reasons are positions; the evidence above does not settle them either way. The marginal cost of machine cognition is falling toward the cost of tokens, and I would rather build on that curve than against it. And the sharpest thing METR found was that nobody could feel where the tools helped and where they hurt, which argues for building the measurement layer and letting it settle the question, rather than trusting anyone’s intuition, the enthusiast’s or the skeptic’s. An agent-default operation with real verification can find the places where humans should keep producing. An operation run on feel cannot find anything, because feel is what the studies showed failing.

Automating the production breaks the apprenticeship that built the judgment

There is one problem here that nobody has solved, and it sits underneath everything else — where the next generation of senior judgment comes from.

For the whole history of skilled work, you learned the hard 20% by grinding through the routine 80%. The junior lawyer read a thousand contracts to earn the judgment that lets a senior partner skim one and find the problem. The judgment was a by-product of the production, and when the production automates, the by-product goes with it.

The studies already show the mechanism running. Earlier I noted that the support agents who gained most from AI were the least experienced, improving around 30% against a 15% average, and the consulting experiment found the same compression, with below-average performers gaining 43% against 17% for those above the line.³² The support-agent study identified how the leveling works, and it is double-edged. The AI captured the tacit knowledge of the best workers and handed it to everyone else.³ That is wonderful for this quarter’s output and corrosive for the career ladder, because it means juniors now perform well without acquiring the judgment the performance used to require. They are handed the senior’s answer instead of building the senior’s instinct.

The leveling result is also only half the literature, and the other half points the opposite way. Willison’s experience is that “AI tools amplify existing expertise,” meaning the more skill and experience you bring, the faster and better the results you get from the machine.⁷ Osmani reports Google’s internal trial showing seniors gaining slightly more, and reads the outcomes as context-dependent.⁸ So the evidence genuinely splits on who gains most, and the split does not need resolving here, because both readings run against the junior. If AI levels, the junior performs without learning. If AI amplifies expertise, the junior has no expertise to amplify and inherits the verification burden anyway. Under either reading, the rung they used to stand on is going.

And the rung is measurably thinning. An analysis of large-scale payroll data found that workers aged 22 to 25 in occupations exposed to AI saw a 16% relative decline in employment, while experienced workers in the same occupations held steady.¹⁰ The decline concentrated in jobs where AI automates the work outright; where it only assists a person, the effect wasn’t there. And the adjustment came through hiring, with firms bringing on fewer junior workers while leaving the pay of those they kept roughly intact.¹⁰ The authors are explicit that this is early evidence, canaries rather than a verdict.¹⁰ The direction, though, is hard to miss. The entry rung is eroding, and it is eroding for exactly the people who used to climb it into judgment.

So what replaces it? From here on I am giving my own position, because the research runs out before this point. I think the new entry rung is the verification and judgment work itself, done under supervision. Juniors enter as apprentice curators and apprentice handlers of escalations. They learn the domain by maintaining the knowledge the agents draw on, which forces them to learn what is true and why, and by working the escalations the agents kick up under a senior’s eye, which is the fastest exposure to the hard 20% there is, because escalations are the hard 20% by definition. The function that makes the AI smart is the same function that makes the junior smart.

The curation half of that deserves its machinery, because without it “maintaining the knowledge” is a slogan. The knowledge the agents draw on is expert judgment encoded on purpose, and it passes through a validation gate before an agent can serve it, because a wrong fact served confidently is the failure mode. It reaches the rented model through retrieval at query time rather than retraining; retrieval is what lets you own the knowledge while renting the brains. And the curation function stays separate from the live escalation queue, because fusing the two turns your knowledge engine into your queue. That is the apprentice’s actual desk: validated knowledge maintained on one side, supervised escalations worked on the other, a senior owning the gate.

There is a structural reason to design the role this way, beyond the humane one. A small senior layer cannot supply exploration by itself; it is too few people, too committed to the current way of doing things, too busy. Exploration has to be somebody’s actual job, and the junior, with less to defend and more to learn, is the right somebody. So the role should write exploring the frontier into the junior’s assignment — finding where the agents fail, red-teaming the outputs, testing the edges — and treat the exploring as the work itself rather than a reward for finishing it.

I want to be honest about the load this position bears. It rests on a bet that curating knowledge and handling supervised escalations builds the same hard judgment a decade of production used to build. The logic is the old learning-by-doing argument. You acquire the skill by doing the thing, and escalations are a concentrated dose of the thing. But it is a hypothesis, and we do not yet have the cohort that came up this way to point to. If the bet is wrong — if judgment genuinely required the volume of routine reps and there is no shortcut through it — then we are hollowing out the path to senior judgment faster than we are building a new one, and the verification layer the whole AI-native model depends on runs short of people who can staff it. That is the real risk in all of this, and it is the one I watch.

What remains for the person is direction, verification, and ownership

Strip the role down and this is what remains. The AI-native individual contributor sets the intent, routes the work, holds a verification layer that catches confident-wrong output before it ships, runs the continuous evaluation that keeps quality drift visible, and owns the result. At the desk this takes two working shapes, and Mollick’s names for them have stuck: the Centaur keeps a clear line between person and machine, delegating whole tasks that sit inside the frontier and keeping the rest, while the Cyborg blends the two, intertwining effort with the AI and alternating control mid-task.⁵ Both run on the same map of where the frontier sits, and the map is earned through use. The daily disciplines underneath are the old engineering virtues run harder. Willison’s working list starts with automated testing and planning before generation and runs through documentation, a review culture, and manual QA, the habits that make machine output checkable at the speed it now arrives.⁷

One test tells you whether any of this is real. A human in the loop beats AI alone only if the human can actually catch the machine’s false positives and false negatives; a reviewer who cannot is theatre, a loop in name only. Everything above — the independent verification layer, the continuous evaluation, the apprenticeship that grows people who can staff both — exists to keep that test passing. Everyone can rent the same models. The judgment about when to trust them has to be built, and the honest open question is whether the next generation can still build it once the production work that used to supply it is gone. I have a wager on how. I do not have proof. Anyone who tells you they do is selling the confident-wrong answer this whole discipline exists to catch.

Sources

1. Sida Peng, Eirini Kalliamvakou, Peter Cihon, Mert Demirer, "The Impact of AI on Developer Productivity: Evidence from GitHub Copilot," arXiv:2302.06590 (2023). Randomized trial; the treatment group completing a JavaScript HTTP server finished 55.8% faster than the control group. https://arxiv.org/abs/2302.06590
2. Fabrizio Dell'Acqua, Edward McFowland III, Ethan Mollick, et al., "Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of AI on Knowledge Worker Productivity and Quality," Harvard Business School Working Paper 24-013 (with Boston Consulting Group), 2023; published in Organization Science, 2026. 758 consultants; inside-frontier gains of +12.2% tasks, 25.1% faster, >40% quality; below-average performers +43% vs above-average +17%; outside the frontier, AI users 19 percentage points less likely to be correct. https://mitsloan.mit.edu/sites/default/files/2023-10/SSRN-id4573321.pdf
3. Erik Brynjolfsson, Danielle Li, Lindsey R. Raymond, "Generative AI at Work," Quarterly Journal of Economics 140(2), 2025 (NBER WP 31161, 2023). 5,172 customer-support agents; +15% issues resolved per hour on average, ~30% for less-experienced workers; mechanism is dissemination of top performers' tacit knowledge. https://danielle.li/assets/docs/GenerativeAIatWork.pdf
4. Joel Becker, Nate Rush, Beth Barnes, David Rein, "Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity," METR; arXiv:2507.09089 (2025). 16 experienced developers, 246 tasks in mature repos; AI allowed increased completion time by 19%; developers forecast a 24% speedup and post-hoc estimated a 20% speedup; ML and economics experts forecast 38% and 39% speedups. METR announced (Feb 2026) a revision of the experiment design; this does not retract the finding but should be noted. https://arxiv.org/abs/2507.09089
5. Ethan Mollick, "Centaurs and Cyborgs on the Jagged Frontier," One Useful Thing, Sept 16 2023. The "jagged frontier": tasks of near-identical apparent difficulty fall on opposite sides of an invisible boundary of AI capability; effective use requires learning where that boundary is through sustained practice. Also names the two effective working modes: Centaurs, who keep "a clear line between person and machine" with a strategic division of labor, and Cyborgs, who "blend machine and person, integrating the two deeply," alternating control mid-task. https://www.oneusefulthing.org/p/centaurs-and-cyborgs-on-the-jagged
6. Matthias Hümmer, Franziska Durner, Theophile Shyiramunda, Michelle J. Cummings-Koether, "AI, Metacognition, and the Verification Bottleneck: A Three-Wave Longitudinal Study of Human Problem-Solving," arXiv:2601.17055 (2026). Pilot study (Wave 1 n=21, Wave 2 n=36, Wave 3 n=23, academic setting): 73.9% AI reliance on difficult tasks; accuracy declining 95.2% → 81.0% → 66.7% → 47.8% with complexity; belief-performance gap of +34.6 percentage points on the most complex problems. Small academic pilot; magnitudes indicative, not workplace estimates. https://arxiv.org/abs/2601.17055
7. Simon Willison, "Vibe engineering," simonwillison.net, Oct 7 2025. Defines "vibe engineering" as professionals accelerating with LLMs "while staying proudly and confidently accountable for the software they produce"; reviewing, testing, and understanding AI output as the distinguishing posture; warns that fast production makes review the bottleneck. Also the source of "AI tools amplify existing expertise. The more skills and experience you have as a software engineer the faster and better the results you can get from working with LLMs and coding agents," and of the recommended disciplines: automated testing (particularly test-first), planning in advance, comprehensive documentation, good version control, a culture of code review, and manual QA. https://simonwillison.net/2025/Oct/7/vibe-engineering/
8. Addy Osmani, "The reality of AI-Assisted software engineering productivity," Substack, Aug 16 2025. Synthesizes the evidence: real gains ≈20-30%, not 10x; "almost-correct" output the dominant time sink (cited by 66% of developers as their biggest frustration); PR review times +91% in one industry report (the Faros "AI Productivity Paradox"), individual speedups not aggregating to team velocity. Also notes Google's internal trial showing seniors gaining slightly more — outcomes are context-dependent. https://addyo.substack.com/p/the-reality-of-ai-assisted-software
9. Anthropic (Engineering), "How we built our multi-agent research system," June 2025. Orchestrator-worker pattern: a lead agent decomposes the query, delegates to parallel subagents with clear objectives and boundaries, and synthesizes; output quality depends on explicit evaluation — an LLM-judge rubric plus human evaluation for edge cases. https://www.anthropic.com/engineering/multi-agent-research-system
10. Erik Brynjolfsson, Bharat Chandar, Ruyu Chen, "Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence," Stanford Digital Economy Lab, Nov 13 2025. ADP payroll data: workers aged 22-25 in AI-exposed occupations saw a 16% relative employment decline while experienced workers stayed stable; concentrated in automating (not augmenting) uses; adjustment via employment, not wages. Secondary coverage cites a related "13% decline in entry-level hiring"; the paper's headline is the 16% relative employment figure. Authors stress this is early evidence. https://digitaleconomy.stanford.edu/app/uploads/2025/11/CanariesintheCoalMine_Nov25.pdf
11. Microsoft (Work Trend Index), "2026 Work Trend Index Annual Report: Agents, Human Agency, and the Opportunity for Every Organization," May 5 2026. Survey of 20,000 workers using AI across 10 countries plus Microsoft 365 telemetry; 86% of AI users say they treat AI output as "a starting point, not a final answer"; Frontier Firm teams discuss quality standards for AI-assisted work at 54% vs 29% for other teams; organizational factors account for roughly 2x the AI impact of individual effort alone. Vendor research with Microsoft-defined constructs and self-reported data; directional framing, not causal evidence. https://www.microsoft.com/en-us/worklab/work-trend-index/agents-human-agency-and-the-opportunity-for-every-organization