The Agentic Adoption Ladder

There’s a particular kind of demo doing the rounds at the moment. An agent picks up a Jira ticket, writes a plan, opens a PR, and someone in the room leans in (quietly, like they’ve spotted a bargain) and starts talking about how few people will need to touch the backlog.

I’ve been in that room. The healthy instinct underneath the excitement is usually better than the wording: fewer humans having to watch tickets, freeing them up to focus on owning their systems. I’ve also been the person who then spent the next fortnight babysitting almost-right diffs, watching token spend climb, and discovering that “operated” still meant a human glued to every merge that mattered. The pitch is Stage 3. The receipt is Stage 2 with a worse burn rate, and more watching, not less.

What we’re arguing about isn’t whether agents belong in software delivery. They already do. It’s which rung of the ladder you’re actually on, and whether the next one is worth climbing yet. Most organisations itching for Stage 3 are better served deepening Stage 2 until the economics clear.

The ladder

Four stages. The labels matter, because vendors will happily blur them into one word: “agentic.”

1. AI-assisted engineering. You still own the ticket. On a Monday morning that looks like opening a C# service change yourself, writing the FluentValidation rules, and letting the agent fill the boilerplate, suggest a test name, or review the diff in chat. The agent is a sidekick. It is not running the work.

2. AI-directed engineering. You turn the business requirement into a brief: acceptance criteria, constraints, “done.” The AI implements under that direction. You still decide whether the PR is mergeable. Same Monday, different shape: the issue says “reject malformed betting-slip payloads with a stable error contract,” you write that as the prompt spec, the model produces the validator and tests, and you keep the merge judgment.

3. AI-operated engineering. Agents triage the backlog, engineer, review, and deploy. Humans step in only for risk and uncertainty, and still own the domain.

4. AI-native. Agents own the delivery loop end to end. I’m parking this one. AI-native organisations are mostly a fairy tale today, which is why we’re not dwelling on it.

The Stage 2 vs Stage 3 boundary, in one line: Stage 2 means the human owns ticket intent and merge judgment; Stage 3 means the agent owns triage and the change loop, with humans only at risk or uncertainty.

Most organisations I talk to are sitting on Stage 2 (or in a messy Stage 1 to 2 blend) and staring hungrily at Stage 3. Stack Overflow’s 2025 survey puts AI tool use almost everywhere (84% use or plan to), but agents themselves are not mainstream: over half of respondents either don’t use them or stick to simpler tools, and only about 14% use agents daily.¹ DORA and Thoughtworks found something sharper still. Among thousands of practitioners, 61% never use agentic mode.² Menlo’s enterprise cut puts true agents at roughly 16% of deployments; copilots still dominate the spend.³ Gartner’s agentic Hype Cycle is blunt: only 17% of organisations have deployed AI agents at all, and fully autonomous agents are “not ready for the majority of enterprise use cases.”⁴

So the ladder is real. The crowd is clustered on the lower rungs. The marketing is already selling the top.

Why Stage 3 seduces

Stage 3 is the “operated fantasy.” Fewer humans having to watch tickets. More time for engineers to own their systems. Agents as digital juniors who never sleep. That’s the version worth wanting.

There’s a worse version of the same pitch. Leadership hears “operated” and pictures headcount leverage without having to say the word out loud. Same demo; different conclusion.

I’ve gotten carried away, even in my personal capacity. We’ve spent a decade drowning in process theatre (Jira states, review queues, deploy rituals), and the idea that an agent might eat the watching is intoxicating. I’ve chased versions of it myself: more autonomy in the loop, less friction on the boring tickets, a quieter Monday spent on the systems that actually matter.

The pathology isn’t wanting to stop babysitting the queue. It’s mistaking a ticket-to-PR demo for a ticket-to-production operating model, or worse, for a headcount plan. Those are not the same animal, and the gap between them is where the money (and the attention) goes.

Why Stage 3 is cost-ineffective now

Token prices fell. Enterprise LLM spend still tripled into end-2025.³ FinOps teams are not imagining this: 93% reported exceeding AI budgets, and roughly one in five had to constrain use because of OpEx.⁵ Agentic work isn’t chat with a hat on. Long context and tool loops burn on the order of ~1,000× the tokens of a normal reasoning exchange.⁵ McKinsey’s tokenomics cut is the one that should sit on every architecture review slide: about 60% of agentic task cost is refinement (check, repair, re-verify), not the first answer.⁵ Same task, ~30× cost variance. Average-cost budgets quietly die.

Then there’s the human tax that Stage 3 brochures forget to print. Stack Overflow: “almost right but not quite” is still the dominant complaint (66%); 45% say debugging AI output takes more time.¹ DORA keeps finding the awkward pairing: AI can raise throughput and raise delivery instability at the same time.⁶ METR’s early-2025 RCT on experienced open-source developers is the cold shower: people expected to be ~24% faster with AI and were 19% slower.⁷ On agent-authored fix PRs in the AIDEV-POP study, more than a quarter closed unmerged; Copilot and Devin merge rates sat around 42-43%.⁸ Answer.AI’s independent Devin trial landed 3 successes out of 20.⁹

None of that says agents are useless. It says unattended operated loops are expensive, unreliable, and still gated by humans who own the blast radius.

The honest contradiction

Well-capitalised teams are running Stage-3-shaped flows. Stripe’s Minions ship over a thousand PRs a week with no human-written code, and still require human review, ban production and user-data work, and sit on years of platform investment (blueprints, millions of tests, serious compute).¹⁰ OpenAI’s harness path gets close to zero manually written lines only after a mountain of AGENTS.md, linters, and agent-to-agent review. GitHub’s coding agent has produced a huge volume of draft PRs; “draft” and “human review” are doing the heavy lifting, not auto-deploy. Cognition’s own Devin numbers improved over the year, and a large share of agent PRs still don’t land cleanly without humans in the loop.¹¹

So yes: Stage-3-shaped ticket-to-PR exists. It is not a cheap SaaS flip. It is a platform play with mandatory human gates. Treating Stripe’s harness as proof that your org should “go operated” next quarter is how you buy autonomy theatre at FinOps prices.

Stage 3 isn’t wrong forever. It’s wrong at today’s prices for typical enterprises. The grown-up move is to stay on Stage 2 and deepen it until the economics clear.

Deepen Stage 2: Monday practices

If directed engineering is where you actually are, treat it like a craft, not a waiting room for autonomy.

Own the intent before you prompt. Acceptance criteria, constraints, and “done” belong in the issue, not in a vibes-based chat. Issue quality is still the strongest predictor of whether the AI output is usable. You’re not typing wishes; you’re writing a brief. On a .NET service that might mean the ticket already states the error codes, the FluentValidation rules that must stay stable for downstream clients, and which integration tests are in scope before anyone opens the agent.

Fortify the system the AI will amplify. DORA’s uncomfortable truth is that AI magnifies whatever delivery system you already have.² Thin tests, sluggish feedback, mushy ownership: agentic tooling will make that louder, not kinder. Platforms before agents: testing, version control, fast feedback loops.

Trust-but-verify as policy, not politeness. Around 30% of DORA respondents report little or no trust in AI-generated code; overall accuracy distrust still outruns trust in the Stack Overflow numbers.¹² Train people to guide, evaluate, and validate, not to rubber-stamp the first plausible diff.

Harness the repo. Short, sharp agent instructions (think ~100 lines, not a novel), progressive docs, and mechanical invariants enforced by linters that tell the agent how to fix itself. This is the boring half of what Stripe and OpenAI actually bought with their “operated” results: not magic models, harness engineering. Steal that part. A useful AGENTS.md for a C# codebase might be a page that names the solution layout, bans async void, points at the shared validation library, and tells the agent which test project to run. Leave the fleet-sized fantasy on the shelf until the unit economics stop looking like a fire.

Instrument what good directed work looks like now (cycle time on AI-touched tickets, rework rate, review burden, spend per merged change) so you have a baseline before anyone proposes a Stage 3 pilot with a slide deck and a dream.

Ready for Stage 3 later: signals, not a timeline

I’m not giving you a date. Dates are how vendors sell ladders.

You’re closer when:

• Cost per successful merged change is stable enough that Finance isn’t inventing new swear words each month.
• Agent triage doesn’t need a human shadow on every ticket just to stop nonsense entering the queue.
• Review burden isn’t eating the leverage the agents were supposed to create.
• Your harness (tests, rules, permissions, blast-radius limits) already treats “almost right” as a first-class failure mode.

Until those signals show up, Stage 3 is optional cosplay. Stage 2 with discipline is the operating model.

Don’t buy the operated fantasy at today’s prices

The ladder is useful. Climbing it on vibes is not.

Most of us are at Stage 2 and itching for Stage 3 because Stage 3 sounds like fewer humans watching tickets, freeing people up to own their systems. At current token economics, refinement overhead, and merge reality, jumping for that itch is a cost trap with good lighting. You often buy more watching, not less. Deepen directed engineering. Own the domain. Build the harness. Let operated-agent economics clear before you rent the fantasy.

If you’ve already tried to jump early and got the bill, you’re not alone.

Sources