Field note · Agentic coding · June 2026
Loop engineering is control-system design
The next step after prompt engineering is not “let the agent run forever.” It is designing a bounded system that finds work, delegates it, verifies outcomes, records state and decides what happens next.
From prompts to loops
A prompt asks for one result. A loop continually discovers or receives work, selects an agent, provides context, checks the output and updates persistent state. The human moves from steering every turn to engineering the operating model.
What a serious loop needs
A production loop needs a source of work, isolated workspaces, reusable skills, connectors, specialist sub-agents and memory outside the conversation. It also needs termination criteria, budgets, concurrency limits, escalation and a way to distinguish progress from repeated activity.
Where software engineering returns
Loop engineering depends on queues, leases, idempotency, state machines, schedulers, tests, telemetry and recovery. The agent may generate code, but conventional engineering determines whether the loop is safe, economical and understandable.
How I would apply it
For a delivery platform, one loop can triage requirements, another can implement a vertical slice and a separate evaluator can test acceptance criteria. Changes only advance when deterministic checks pass; uncertain architectural decisions return to a human reviewer.
Executive takeaway
The value is not fewer humans typing prompts. The value is repeatable throughput with explicit controls. Measure completed business outcomes, rework, cost per accepted change and escaped defects, not agent activity.