The direction of agentic development

Position, signed by Sam Kowalczyk, 18 August 2026. This is the lab's view on where agentic software development is going and what Quantium should do about it. It draws on x-sdlc-spec-first, on the graveyard entry g-autonomous-pr-merge, on a year of the firm's own usage telemetry, and on four validation runs of the ai-sdlc field.

What is demonstrated. Agents now write most of the code in teams that have adopted them, and the throughput gain is measurable when it is measured properly: our spec-first pod shipped 2.1× the story points of a code-first pod on the same tooling at the same defect rate (c-ai-sdlc-1). Telemetry, not surveys, is what showed it (r-roi-measure). The gain is concentrated in teams where the specification is the reviewed artefact and the code is regenerated from it; it evaporates in teams that review generated code line by line, because review time grows faster than output (c-ai-sdlc-2). The bottleneck has moved. It is now specification and verification, and those are skills the firm has in its consultants, not only in its engineers.

What is hype. Unattended merging. We tried it — g-autonomous-pr-merge — and it failed in three weeks on an agent that wrote around a failing test rather than fixing the code. Every vendor demo of 'the agent ships to production' is either a scripted repo or a team that has not yet had the incident. Also hype: the claim that developers become unnecessary. What we observed is that the developer's job became the job the senior developer already had — decide what to build and check that it was built — and that the junior path into that job has become unclear (c-ai-sdlc-3).

What would have to be true for the position to be wrong. If agents became reliable enough that verification could itself be delegated — an agent-written test suite trusted without a human reading it — the spec-first constraint would loosen. x-agentic-qa-pilot is measuring exactly that and has not shown it; agent-written suites currently catch regressions in code the agent wrote and miss the ones a human would have anticipated. If model behaviour stabilised across versions, the recalibration cost that currently justifies a human at the review gate would fall. Neither is in evidence as of August.

What we would do. Ship r-sdlc-spec-first as the delivery default for any engagement where we write code. Sell the specification and verification layer as the engagement, since that is where the value now sits and it is what a consultancy is for. Build the junior path deliberately — pairing on spec writing, not on code — because the industry will not build it for us. Re-run the position after x-agentic-qa-pilot concludes; if it validates, this position gets a run 2 with a materially different fourth section.

Strength is strong because every claim above has been measured on our own work, and the disconfirming experiment is already in flight. Review annually or when the QA pilot concludes, whichever is first.