A post by Marco Vargas at Gorilla Logic started this one. His take on the AI-First SDLC named something I'd been watching happen and hadn't quite put into words.
Most developers didn't notice the shift when it first happened. One day you were writing a line; the next, a tool was finishing it. Code completion (Copilot, tab complete) kept the framing the same. You were still the programmer. The tool was a faster typewriter.
The second shift was harder to date. You stopped writing functions and started describing them. Conversational AI turned the editor into a dialogue. You were still directing every task, but the medium changed. AI-assisted development: powerful, but still organized around you as the author.
Marco puts a name to the third shift. In his AI-First SDLC, the developer stops being the author and becomes the orchestrator. Agents take on roles—product owner, architect, developer, tech lead, QA—and the human governs the process rather than executing it. The shift isn't about better tools. It's about a different relationship to the machine.
Marco calls the moment of recognition "sharp," and it is. The progression from stage two to stage three isn't incremental; it's a reframe. The tools don't change much. The mental model does.
Why It Works Solo
Marco's model works because the orchestrator and the context-holder are the same person. One brain holds the governance rulebook, the intent behind the requirements, and the judgment on whether the agent's output is actually correct. The human gates between sprint phases aren't a bottleneck—they're the entire quality system.
When context fits in one head, coordination overhead is zero. There's no ambiguity in handoffs, no governance layer to route through. It works because the scope is contained.
Why Enterprise Is Different
Scale breaks exactly what makes it work. Governance becomes committee-owned and drift-prone. Intent gets distributed across a dozen people with a dozen different read-outs. The human review gate, lightweight when one person holds all the context, becomes the new bottleneck when verification has to be spread across a team that didn't write the code and can't assess it at the rate it arrives.
That's the dynamic The Throughput Trap traced in 2026 data: individual output up 66%, organizational delivery flat, the constraint having moved from writing to verifying. Marco is right about the destination. Getting there with a hundred engineers is a different map.