autonomic software engineering

The autonomic nervous system regulates the body without conscious attention. It is not autonomous — autonomy implies a goal of its own. The autonomic system serves a goal that consciousness sets; it adjusts heart rate, digestion, temperature without asking. The etymology is load-bearing.

Kephart and Chess (2003) named autonomic computing for the analogous property in systems: the runtime manages itself, with the operator setting policy and the system handling the rest. MAPE-K — Monitor, Analyze, Plan, Execute, Knowledge — was the architecture. The hard part, the part that prevented anyone from actually building an autonomic system, was the Analyze stage. You needed a model of intent, and nobody could write one down.

That constraint has moved. The model of intent is now something a model can produce. The hard part is no longer generating a change. It is deciding, without a human, whether the change was an improvement. Every architecture that fails this fails the same way: the thing that proposes the change also decides whether it worked. The loop then optimises the judge. This happens quietly, produces a rising number, and is usually discovered six weeks later.

Autonomic software engineering is the practice that points the MAPE-K loop at the engineering process itself. The codebase is the managed subject. Sensors observe the consequence of a change. The proposer emits a patch. The applier lands it. The scorer runs the eval. The verdict plane decides pass, fail, kill, harden, or pivot. The cycle converges against an eval, not against a reviewer.

What it is not

It is not agentic engineering (Karpathy, 2026) — which is about how a human directs agents. Agentic engineering describes the operator. Autonomic software engineering describes the change path.

It is not an autonomous software engineer — a role description, not a system property. Different subject, no conflict.

It is not self-modifying AI in the sci-fi sense. The system has a fixed target (the eval) and converges against it. The system is not improving itself in any unbounded sense; it is closing a loop on a defined subject.

Position

There is 2004-era prior art applying autonomic concepts to software development operations. It predates any component capable of delivering it and has no practitioner footprint. Citing it pre-empts the objection at no cost.

The IEEE / ACM Automated Software Engineering conference has used “ASE” since 1986. The acronym is therefore not available. The full term is the name; the practice is what it is.

See also

• closure-test — the four-condition rubric that decides whether a system qualifies
• two-plane-loop — the reference architecture for separating the change from the decision