Eighteen months ago, AI coding agents were conference demos. Today they sit inside our delivery pipeline on most projects. This is an honest account of what changed in our own engineering practice over the past year — measured against sprint data, not marketing enthusiasm.
From demo to daily tooling
The shift wasn't a single moment. Through 2025, agentic tools crossed three practical thresholds: they became reliable enough to run against real codebases without constant supervision, cheap enough to run on every ticket rather than special occasions, and — most importantly — auditable enough that our code review process could treat their output like any other engineer's pull request.
Across Mtoag's delivery teams, roughly 60% of production tickets now involve an agent at some stage — scaffolding, test generation, migration work, or first-draft implementation. That number was near zero at the start of 2025.
Where agents genuinely accelerate work
Our sprint data shows consistent, measurable acceleration in four areas:
- Boilerplate and scaffolding — new modules, CRUD endpoints, and configuration land 3–4× faster with no quality difference after review.
- Test coverage — agents write thorough unit tests for existing code faster than any human wants to. Coverage on legacy modules we maintain has risen from ~45% to ~80% at near-zero marginal cost.
- Migrations and upgrades — framework version bumps and API migrations, the least-loved engineering work, are now largely agent-driven with human spot-checks.
- Documentation — generated first drafts mean documentation actually exists, which was not reliably true before.
Notice what's on that list: work with clear correctness criteria. Where "done" is checkable, agents excel.
The review debt problem
The uncomfortable finding: agents move effort rather than removing it. Code that took a day to write and an hour to review now takes an hour to generate and half a day to review properly. If your review discipline is weak, agents will happily fill your codebase with plausible-looking code that nobody on the team truly understands.
We've responded structurally: agent-generated changes are labelled in every pull request, review time is budgeted explicitly in sprint planning, and no agent output merges without a named engineer taking ownership of it — the same accountability rule we apply to human-written code.
Scoping AI-assisted projects
For clients, the practical consequence is that project economics have shifted unevenly. Well-specified builds with conventional architectures are genuinely 20–30% faster to deliver than two years ago — and our fixed-price quotes reflect that. Novel, exploratory work is not much faster at all, because the bottleneck there was never typing speed; it's decision-making, and that remains human.
Be suspicious of any agency promising uniform AI-driven cost reductions across all project types. The gains are real but concentrated, and a supplier who can't tell you where they materialise probably hasn't measured them.
What stays human
Architecture decisions. Security review. Anything touching payments, personal data, or regulatory surface. Client communication about trade-offs. And final accountability — a named engineer owns every line that ships, however it was produced.
The teams winning with AI agents in 2026 are not the ones that adopted the most tools. They're the ones that adapted their engineering discipline fastest — because agents amplify whatever process they're dropped into, good or bad.
Scoping an AI-assisted project?
We'll tell you honestly where AI accelerates your build — and where it won't.
