Fast delivery isn't magic. It's more QA, not less.

Every few weeks someone asks a version of the same question: *how are you shipping production software this fast?* The honest answer is boring. The coding part got faster. Everything around the coding part got more important.

A recent CIO piece on agentic engineering workflows puts it well: autonomous coding agents are moving into day-to-day pipelines, but the wins only show up when teams redesign process, review, and infra around them. That matches our experience. If you drop agents into a shop that already had loose QA and manual deploys, you just get bad code faster.

What actually happens on a build

A typical Arkitekt project looks like this under the hood:

- Discovery, by a human. We map the workflow, the data, and the edge cases before anyone writes a line of code. Agents are bad at inferring what a business actually does. People are bad at explaining it. That gap is where projects die. - Scoped generation. Agents write in small, testable slices against a spec we wrote. Not "build me a CRM." More like "given this schema, write the sync job with these three failure modes handled." - Tests written alongside, not after. Sometimes the agent writes them. Sometimes we do. Either way, the tests are the contract. If they pass, the slice ships. If they don't, the agent iterates until they do or a human takes over. - Human review at the seams. Code review still happens, but the focus has shifted. We spend less time on syntax and more time on "does this behave correctly when the third-party API returns a 429 at 2am." - Managed deploys. Every project ships on infra we operate. Same deploy pipeline, same monitoring, same on-call. The client doesn't get a repo and a shrug.

What breaks first

The Harness CTO commentary on how AI is changing software delivery in 2026 argues that QA, release management, and infra are the parts that break first when the build phase speeds up. That's been true for us. When you can generate a working feature in an afternoon, your bottleneck is no longer typing. It's verifying, deploying, and monitoring.

So we've invested there. Evaluation harnesses for the parts of the system where an agent has judgment. Guardrails on what the agent can touch in a production database. Rollback that works in one click. A recent arXiv guide on production-grade agentic workflows catalogs the same patterns — evaluation, guardrails, human oversight — as the difference between a prototype and something you can actually run.

The implication

"Days to deliver" is real, but it's not because the agent is a wizard. It's because the surrounding discipline is tighter than it used to be. If someone's pitching you AI-built software with less QA and less infra work, be skeptical. In our experience, it's the opposite. The code gets cheaper. The care around it gets more expensive, and more worth it.