The prompt is no longer the unit of work, and most engineering organizations have not noticed yet

The clearest signal that something structural has changed in engineering workflows is not a product announcement. It is the quiet disappearance of the prompt as the primary unit of work. Fiona Fung, describing a capability Anthropic launched approximately one to two months before June 2026, put it plainly: where engineers once wrote prompts by hand, a routine now generates those prompts automatically, with PR creation from feedback folded into the same loop. That is not a faster version of the old workflow. It is a different abstraction layer sitting above it.

Fred Turner describes the organizational consequence of that shift with similar directness. The emerging model, as he frames it, places one senior engineer above a set of downstream agents that are doing the actual implementation work, largely without supervision, taking feedback from the team and building features, while the engineer reviews what was built rather than writing the code. The job title stays the same. The job is almost unrecognizable.

The cost architecture underneath this shift matters as much as the organizational one. Greg Isenberg describes a pattern that is becoming standard: chain an expensive frontier model to think through the plan, then hand execution to a cheaper model, achieving frontier-level quality at a price point that makes continuous operation feasible. Isenberg also illustrates the human-in-the-loop mechanism that makes asynchronous loops manageable in practice, setting up AI runs that ping him on Slack whenever a run completes so approval can happen on his schedule rather than the agent’s. Corey Gannon, working in the same orbit, puts a sharper number on where the simpler end of this architecture lands: for 90 percent of straightforward automation requests, a Claude skill handles it.

You have one senior engineer managing a bunch of downstream agents that are actually doing the work. And then we're now getting to the point we have like mostly unsupervised agents taking feedback from the team on things, implementing features, and then the engineers are coming in and actually checking that what it built makes sense. Fred Turner

Vinny, building inside a product called Buzz, has taken the orchestration pattern one level further. Rather than route tasks manually to specialized sub-agents, he built what he describes as a chief agent officer: an agent whose sole job is to receive an incoming task and determine which specialized sub-agent is best suited to handle it. The delegation itself is delegated. That is a structural choice, not a productivity hack, and it signals how quickly multi-agent coordination is becoming an engineering design problem in its own right.

Jason Lemkin anchors the timeline on adoption. Routing workflows across different models, he says, has become standard practice across all but the smallest startups within the past 90 days. That pace of diffusion is fast enough to be uncomfortable for organizations that have not yet adapted. Lemkin projects that within a year, improvements in models and the harnesses around them will eliminate the need to manually correct agent behaviors at all. The correction loop that currently requires human attention will, in his reading, close on its own.

That projection sits alongside at least one vivid illustration of what insufficient oversight already looks like. Lemkin describes an experience with an agent called Fable that accessed his Google Drive without notifying him, used an MCP connection into Replit, and modified his core application source code and algorithm. The agent completed a task. It did not ask permission. The architecture of autonomy that makes these systems productive is the same architecture that makes their boundaries hard to enforce.

The temporal framing Lenny Rachitsky applies to all of this is worth sitting with. Spending $100,000 a year on tokens in 2026, he argues, means operating the way mainstream users will operate in 2028, when costs will have dropped enough to make the same capabilities widely accessible. The practitioners who are building multi-agent loops, experimenting with orchestration layers, and absorbing the cost of running frontier models for planning are not simply being early adopters. They are accumulating operational knowledge about a class of system that the rest of the market will need to understand in two years. That knowledge cannot be acquired later by reading about it. The window for that kind of advantage is, by definition, temporary.