alphasignal loop engineering loopmaxxing

Why this is in the vault

AlphaSignal's Sunday Deep Dive by Ben Dickson (TechCrunch/VentureBeat) crystallizes the "loop engineering" paradigm that Boris Cherny (Anthropic, Claude Code) and Peter Steinberger (OpenClaw) are now calling the successor to prompt engineering. The piece defines the five loop primitives, names the failure mode ("loopmaxxing" = open-ended while-true iteration without exit conditions), and lays out the pragmatic mitigation path. Directly relevant because RDCO's harness is itself a production loop — automations, worktrees, skills, sub-agents, memory — and the anti-patterns described are live risks.

⚠️ Sponsorship

Weights & Biases (wandb.ai) sponsored this issue with a guide on connecting evaluations, traces, human review, and approval workflows for agent deployment readiness. Content is editorially independent; sponsor plug is a sidebar, not embedded in the analysis.

Issue contents

Sunday Deep Dive: "All about loop engineering (including the pitfalls)"

• What is loop engineering? Shifts LLMs from call-and-response to active participants in an event loop. Developer specifies a verifiable goal; the loop observes state, acts, checks result, decides to continue/retry/stop. Five primitives per Addy Osmani (Google): Automations (triggers), Worktrees (isolated branch environments), Skills & external tools (markdown guidelines + MCP integrations), Sub-agents (specialized models with evaluator/executor split), Memory (external tracking because context windows flush).
• The loopmaxxing trap: Open-ended "while(true)" cycles with fuzzy goals cause infinite drift. Agents reviewing their own flawed sub-agents reinforce mistakes. Token burn is structural — retries, failed tool calls, and context reconstruction accumulate. "Comprehension debt" widens between repo state and engineer understanding.
• Pragmatic path: Write strict control loops — deterministic code for execution, LLM only for dynamic decisions traditional code cannot handle. Start with minimal loops + human verification. Prove manually, then automate incrementally. Separate executor and evaluator agents. Hard caps: max 2–3 retries before graceful failure + human handoff. Trace-logging + progress detection mandatory in production.
• Key quote (Boris Cherny, Anthropic): "I don't prompt Claude anymore. I have loops that are running. They're the ones prompting Claude... My job is to write loops."

Mapping against Ray Data Co

Loopmaxxing exposure: The open-threads-check, morning-prep, and check-board crons are low-risk (bounded scope, human-readable output). Higher risk lives in any multi-step build dispatches without explicit iteration caps. The "max 2–3 retries before graceful failure" rule is not yet codified in RDCO skill dispatch SOPs — worth adding as a hard-exit gate, especially for investing execution scripts and deploy agents.

Comprehension debt: The "agents run loudly and fail quietly" observation maps directly to the no-batched-result-declaration memory and the verification-as-independent-worker pattern already in place. The principle here validates that discipline.

Actionable: Consider adding an explicit max_retries: 3 convention and hard iteration-cap comment block to any skill that dispatches sub-agents in a retry loop (xcode-build-fixer, pipeline patterns, investing deploy scripts).

Related

• [[2026-05-30-alphasignal-solo-vs-team-agent-enterprise-deep-dive]] — agent team architecture and executor/evaluator split at enterprise scale
• [[2026-06-07-alphasignal-async-agents-release-bottleneck]] — async agent patterns and release bottleneck analysis
• [[2026-04-08-better-harness-evals-hill-climbing]] — harness hill-climbing with evals; the systematic loop-tuning complement to this piece
• [[2026-05-12-zach-lloyd-warp-verify-then-build-test-harness-agentic-coding]] — "verify then build" as a deterministic guardrail pattern for agentic coding loops