Loop specifications: the check is the skill, not the prompt
Jun 28 Macedo: defines the external loop spec (trigger, goal, verify, stop, memory); codes 50 Loop Library entries—70% autonomous verify, but triggers and durable memory still mostly manual.
Source: Stop Hand-Holding Your Coding Agent: Engineering the Loops that Replace Step-by-Step Prompting
Sandeco Macedo (IFG) takes the mid-2026 slogan seriously—stop prompting the agent every step; design the loop that prompts it—and gives the practice a reviewable object. The loop specification is an external, bounded, reusable artifact (trigger + goal + verification + stopping rule + on-disk memory) handed to a harness such as Claude Code or Codex. It is neither a programming while nor the harness’s internal perceive–act–observe cycle: the harness is the engine; loop engineering writes the pilot. The paper places loop engineering as a fourth layer over prompt → context → harness, and refuses the stronger headline that loops retire prompting.
The useful engineering surface is a five-level verification ladder (deterministic → rule/schema → delayed field truth → model-as-judge → human checkpoint). Levels 1–2 are the autonomous zone; pretending level 4 is level 1 is named as an anti-pattern beside self-approving makers, specification gaming, and unattended runaways. A hand coding of the public Loop Library (fifty loops) shows practice matured where the discipline says value lives: 70% verify in the autonomous zone, 74% name terminal states, 66% set a verifiable goal—while only 22% use automated triggers, 20% call named reusable skills, and 32% develop persistent memory. Accompanying deliverable: sandeco-loop, an open skill that authors hardened loop specs (triage → interview → anti-pattern hardening → single .md actuation) rather than running them.
Why it matters for code review
Always-on factories that can already prove “done” still stall on the half that removes the human from starting and remembering: schedules, skills, and curated disk state. Review overnight lanes against Macedo’s maturity mismatch—strong gate, weak trigger/memory—before FOMO for idle agents buys another solo retry loop with a self-score. Steady harnessed throughput compounds when the check is external and terminal states are named; peak-session mastery that exits on the implementer’s optimism does not. Pair with Lulla’s 217-loop adoption mine and LoopsBench’s harness-to-loop eval when promoting unattended cadence past a green CI badge.