Skip to content

Knowledge base

The State of AI-Native Engineering report

Augment Code's survey of 219 engineering leaders found that about half their code is AI-generated, yet only 57 changed code review and fewer changed roles, onboarding or measurement. Losing shared understanding of the codebase worried 55%, the top concern and the least addressed.

Written by

Updated:

Evidence 7

Our position. We see the gap this survey measures in most rollouts: they change the codebase and not the team. Agentic engineering is how the codebase, the engineers, the reviews and the non-engineers move together (codebase versus people and process (opens in a new tab)). Worries about AI are not a training problem, because management decides what happens to roles, so they go on the agenda before the tools.

Where DORA and Faros measure what happens to the code, this survey measures what happens to the organization, and finds it has not moved. 219 engineering leaders, managers through CTOs, mostly leading teams under 50 (Augment Code - The State of AI-Native Engineering in 2026 (opens in a new tab)). The sample was recruited through engineering-leadership networks and skews AI-forward, and Augment sells an agent orchestration platform. Read the numbers as directional, from leaders further along than average.

How much code is AI-generated, and do leaders trust it?

Roughly half of all code in these organizations is now AI-generated (mean 48.3%, median 48%). Volume is not confidence: 39% of respondents are worried or very worried about shipping with confidence as agents write more code and approve PRs, and only 12 of 219 are not worried at all. Leaders rate their competitiveness higher looking forward than looking back (3.96/5 against 3.66/5) and simultaneously do not trust what they are shipping. Both are true at once, from the same people.

CircleCI’s telemetry says the same thing from the pipeline (The 2026 State of Software Delivery (opens in a new tab)): across 28 million CI workflows, average throughput rose 59% year over year while the main-branch success rate fell to 70.8%, its lowest in five years. More code is being written, and more of it is breaking.

What worries engineering leaders most?

55% of respondents are concerned or very concerned that humans are losing shared understanding of how the codebase evolves. It is the loudest signal in the dataset, and it is the thing fewest organizations are doing anything about. The survey maps to a three-layer model of system health (technical, cognitive, and intent debt), the mechanics of which are on Quality and Verification.

The behavioral risk under it has a name. Wharton research distinguishes cognitive offloading, which is strategic (you know what you delegated), from cognitive surrender, which is passive: you stop applying your own judgment because the output looks right and checking it is expensive. It reads as efficiency right up to the point where something breaks and nobody can trace the cause.

What have organizations actually changed?

Only the two artifacts closest to the code have moved. Of 219 leaders, 57 formally changed code review and 57 changed specs. Everything else is stalled: measuring productivity 32, engineering operations 24, role definitions 19, onboarding 15.

This is the sharpest finding for anyone running an AI transformation. The hiring bar has already moved (leaders now rank evaluating and verifying AI-generated code, orchestrating agents, and product judgment above traditional coding) but the job descriptions, interview loops, career ladders, and onboarding paths that encode that bar have not. The specs, the ladders, and the definitions of “good” are still written for the old world, for humans and agents both.

The three debts compound in exactly this gap: when the code is less understood and the specs describing what it should do are stale, the technical debt becomes invisible until something breaks.

How do engineers take the shift?

63% of leaders say engineers on their team are voicing concerns that their skills are becoming less relevant. At large teams (201 to 1,000 engineers) it is 89%; at the smallest it is 55%. The anxiety scales with org size. Set against the 19 who changed role definitions and the 15 who changed onboarding, the conclusion is uncomfortable and direct: most managers are absorbing these conversations and then doing nothing. Feelings that get no answer do not disappear, they go underground as passive resistance and quiet disengagement.

Two phrasings from the report’s own team name the shift precisely. Engineers describe moving from “proud builder” to “proud coordinator”. And the aspiration is not “ticket to PR” but “intent to production”.

Headcount, meanwhile, is not collapsing: 81 flat, 68 modestly increasing, 14 significantly increasing, against 30 modestly and 6 significantly reducing. Mid-size orgs move hardest in both directions (61% plan to grow, 22% to reduce). This is job redefinition, not job loss.

What does “AI-native” mean?

Asked to define the term in their own words, 219 leaders produced everything from full organizational redesign to autocomplete with extra steps to “a nonsense marketing term”. Teams are being asked to move toward something with no shared definition. Before running an adoption program, define the target concretely enough that a team can tell whether it hit it.

What does the report recommend?

  • Test for the skill you claim to value. If evaluating AI-generated work is the top hiring priority, the interview loop should assess it.
  • Make specs a living contract, not a document. Spec-driven development fails the way every documentation-first initiative failed, as unrewarded maintenance work. It only holds when agents read from and write back to the spec, updating it when they discover an assumption was wrong.
  • Write down what the team knows. Agents replicate what is already there. Conventions and architectural reasoning that live in people’s heads never reach the agent.
  • Give agents both independence and autonomy. Independence is context, autonomy is the permission to act. An agent with one and not the other is not productive.
  • Make space for the feelings. Take time off the roadmap and ask. The organizations that do keep their best people.

Resources

Research