The Loop Builder's Field Manual

The working field manual for the moment you stop prompting agents one message at a time and start designing the systems that prompt them. Pick the right loop, build one that converges, and run it for hours without blowing up. Inside, on nine dense pages: (1) THE 5-STAGE LOOP every working agent system shares — Discover, Plan, Execute, Verify, Iterate — and the core line that anchors all of it: a reliable loop beats a perfect prompt. (2) THE 3-QUESTION PICKER that makes the architecture choose itself — Control (high → single/sequential; low → hierarchical/swarm), Complexity (single domain → single; predictable → workflow; open-ended → multi-agent), and Budget (tight → single/parallel; and the warning that a multi-agent fleet costs 10–15× the tokens) — plus the default rule: add Skills to a single agent before you build a fleet. (3) THE 6-ARCHITECTURE CHEAT SHEET — single, sequential, parallel, evaluator-optimizer, hierarchical fleet, and collaborative swarm — each with its best-for, its failure mode, and a cost tag, so you stop reaching for a fleet when a single agent with a skill would do. (4) THE 8-STEP CONVERGENCE CHECKLIST with copy-paste code: a deterministic oracle you run 10× to kill flaky checks, a measured baseline run, the minimal stateless bash loop, narrow per-iteration context on a token budget, anti-reward-hacking (read-only tests + a git-diff gate + a cross-model judge), memory on disk (STATUS.md + .loop_state.json), isolation (git worktree → docker --network none --read-only), and brakes + observability (MAX_ITER, budget cap, circuit breaker, heartbeat, JSONL log). (5) THE 4 WAYS LOOPS DIE and the fix for each — Runaway (caps), Silent death (fresh context + heartbeat), Random walk (a deterministic fixpoint check), and Understanding debt (a mandatory human read — the one with no mechanical fix). (6) THE RECEIPTS from Anthropic's blueprint — Coinbase ($226B / 99.99%), Intercom's Fin (86% resolution), Inscribe (20× faster fraud review), Thomson Reuters (3,000+ experts), and the headline result: a multi-agent system beat a single agent by 90.2% (an Opus lead orchestrating Sonnet subagents). The line to remember: pick it, make it converge, then stay the engineer. Architectures and figures are reported as of 2026 — verify before you rely on them.

By submitting, you'll be subscribed to the Hyperautomation AI Report newsletter — one practical AI email per week. Unsubscribe anytime.

Free. No spam. Unsubscribe anytime.