Agentic SDLC and the future of software engineering

In session 5, with Pooria, Soheil, and Omid, we mainly discussed the future of agentic SDLC. I joined from the train while travelling back to London from Scotland, so we also quickly talked about Scotch :D

My main argument was that a traditional SDLC has many steps between gathering requirements and getting something running in production. With sufficiently capable AI agents, perhaps 80% of those steps could eventually disappear or become fully automated. Because of that, I’m fairly pessimistic about the traditional software-engineer role over the longer term.

Pooria pushed back with an example of a migration where Opus 4.5 got stuck and humans had to intervene. My argument was that this is partly a harness and tooling problem: if we build reliable deterministic tools for migrations, testing, validation, and deployment, the AI mainly needs to select and orchestrate those tools rather than solve everything from scratch.

Soheil also brought up the “LLMs Can’t Jump” article I had previously shared. We discussed induction, deduction, and particularly abduction, which seems to remain a major limitation of current AI. Models can work extremely well within existing concepts and systems, but genuinely inventing new concepts based on experience is harder. For now, human experience can help fill that gap, although I think future AGI systems with their own accumulated experience could potentially reduce that limitation too.

Overall, the discussion reinforced my view that software engineering is likely to change dramatically, with much more emphasis on building deterministic tools, harnesses, and environments for agents, while agents increasingly handle the actual implementation work.