continuous-agent-loop

Continuous Agent Loop

Operating discipline for a loop that is already running — as opposed to autonomous-loops, which covers choosing the loop's architecture in the first place. This skill is about the gates that decide whether an iteration's output is good enough to keep, and about recognizing and recovering from a loop that has quietly gone wrong.

Quality gates per iteration

Every iteration of a continuous loop should pass through gates before its output is accepted, not just "the agent said it's done":

1. Build/compile gate — the code compiles and the type checker (or equivalent) is clean.
2. Test gate — the existing test suite passes, plus any regression test written for a bug this iteration fixed (see ai-regression-testing).
3. Acceptance-criteria gate — the specific, concrete criteria defined before the iteration started are met, not a general "does this seem right" judgment made after the fact.

A loop with only an implicit gate ("the agent's summary looked reasonable") has no gate at all — it will accept regressions the agent didn't notice it introduced.

Recognizing failure modes

A continuous loop can look "alive" — consuming compute, producing diffs, committing — while making no real progress. Watch for:

• Loop churn without measurable progress — iteration count climbs but the acceptance-criteria gate never gets closer to passing; each iteration touches different code without converging.
• Repeated retries with the same root cause — the same test fails, the same build error recurs, across iterations that don't look identical on the surface but share a diagnosis. This means the loop is retrying blindly rather than incorporating what the last failure taught it.
• Merge or landing stalls — for loops that merge work (a PR loop or a multi-unit DAG), units repeatedly failing to land cleanly signals a conflict the loop isn't equipped to resolve on its own.
• Cost drift from unbounded escalation — cost per iteration creeping up, usually from repeatedly escalating to a stronger/more expensive model on the same failure instead of fixing the underlying under-specification (see the escalation guidance in agentic-engineering).