Two articles crossed my desk recently that, read together, explain more about the state of AI in software engineering than either does alone.
The first maps the six levels of AI adoption engineers go through – from the Skeptic who dismisses AI after one bad experience, through the Assisted Coder who treats it as autocomplete, up to the AI-Native Engineer who reviews more code than they write, and finally the AI Architect who designs codebases for agents, not just humans.
The second looks at the same story from the top down: the enterprise AI maturity gap. The numbers are sobering.
42%
of companies abandoned generative-AI initiatives in 2025 – up from 17% the year before
2.2
~50%
Here’s what struck me: these are two ladders describing the same climb. And most organizations are failing at the rung where the ladders are supposed to meet.
| Individual adoption · STRV | Individual adoption · STRV 2 |
|---|---|
| L1 Skeptic Dismisses AI after one bad experience | 1 Aware Understands AI exists, no real adoption |
| L2 Assisted Coder Uses AI as glorified autocomplete | 2 Active Pilots underway, mostly disconnected |
| L3 Collaborator Iterates with AI on real tasks | 3 Operational Some AI in production, inconsistent |
| L4
AI-Native Engineer Runs parallel agent sessions, reviews more than writes | 4
Systematic Reproducible, observable, cost-controlled AI systems |
| L5
AI Architect Designs context, guardrails, and verification loops - agents testing agents >$10 | 5
Transformational AI as core business capability |
An organization can’t be “AI-driven” if its engineers are stuck at Level 1, pasting error messages into a browser tab. And an engineer can’t reach Level 4 or 5 if the organization hands them nothing but a chatbot license and a policy document.
Look at what the upper levels of individual adoption actually demand. At Level 4, an AI-Native engineer runs multiple agent sessions in parallel, each working a different task. At Level 5, the work becomes designing context, guardrails, and verification loops – agents testing other agents’ output, automated checks running before code ever reaches a human. As Klacko puts it: the dev job becomes an ops job.
Now ask: where do all those agents run?
Today, mostly on the developer's laptop. They die when the lid closes. They share the developer's credentials and can touch everything on the machine. They produce no durable traces. Every parallel session competes for the same cores.
This is the unglamorous reason the maturity gap persists: organizations approve the tools but never build the runway. The engineer who wants to climb from Level 3 to Level 4 hits a wall that isn’t about skill or mindset – it’s infrastructure.
The Janea piece makes the same point from the enterprise side. The most expensive failure point is the proof-of-concept-to-production gap, and the organizations that cross it treat AI as software engineering first: production-grade environments, reproducibility, observability, cost control. In other words, the boring stuff. Our stuff.
At Incredibuild, we’ve spent more than two decades on a simple premise: developer productivity is an infrastructure problem before it’s anything else. We built our business helping teams at Microsoft, Amazon, Adobe, and HSBC turn waiting time into shipping time by distributing compute at scale.
The agentic era doesn’t change that premise. It amplifies it.
When one engineer supervised one machine, infrastructure shortcuts were survivable. When one engineer supervises ten agents working overnight, the questions become unavoidable:
This is why we built Islo – secure, long-running computers for AI agents. Each agent gets its own isolated environment: a real machine with its own services, not an ephemeral container, with scoped credentials injected at the network edge so secrets never reach the model, snapshots for reproducible environments, and a lifecycle that doesn’t depend on a human keeping a laptop open.
It’s the missing rung between the two ladders: the platform layer that lets individual Level 4-5 workflows become organizational Level 4-5 capability.
This isn’t theoretical for us. Our engineers are experimenting at the frontier Klacko describes as Level 5. One recent internal example: a meta-harness proof of concept in which an agent reads the full execution traces of prior agent runs and rewrites its own harness – going from 0/5 to 5/5 on a task suite in four self-improvement steps.
The entire loop runs on three sandbox primitives: snapshot an environment once, fork it cheaply per task, keep durable logs the proposer can grep through. That experiment works precisely because the infrastructure makes it cheap to be reproducible, parallel, and observable. Which is the whole argument in miniature: the ceiling on your team’s AI maturity isn’t your engineers’ curiosity – it’s whether your platform makes the next level cheap to reach.
Find your team on the individual ladder honestly. Most teams are spread between Levels 1 and 3, whatever the slide deck says. Then stop treating adoption as a training problem.
The jump from Level 0 to 1 is mindset. From 1 to 2 is tooling. But from 3 upward, it’s permission plus platform: low-stakes room to experiment, and infrastructure where parallel agents can run safely without betting the company’s secrets on a laptop’s uptime.
And remember the closing warning from Klacko’s piece, because it’s the right one: speed doesn’t transfer responsibility. The agent writes the code; your name is on the commit. The job of engineering leadership in 2026 is to build the environment where keeping your hands on the wheel is actually possible at speed.
We’ve spent twenty years making fast safe. We intend to spend the next twenty making autonomous safe.
Nir Tzur
Nir Tzur is Senior Vice President of Engineering at Incredibuild, where he leads the platform engineering teams behind build acceleration and agent infrastructure. incredibuild.com
Incredibuild empowers your teams to be productive and focus on innovating.
