StackMap
Subscribe

autoresearch vs loop-engineering

Karpathy's autoresearch loop as an installable skill for Claude Code, OpenCode and Codex: constraint + mechanical metric + autonomous modify→verify→keep/discard iteration. — versus — Reference repo plus npm CLIs (loop-init/audit/cost) for loop engineering: designing scheduled, gated control loops that prompt and orchestrate AI coding agents — Grok, Claude Code, Codex — over time.

The curated verdict

Both operationalize 'the loop is the unit of progress': loop-engineering is the design methodology and CLIs for building gated loops; autoresearch is one specific, proven loop shipped as an installable skill.

autoresearchloop-engineering
Stars5.3k8.0k
Forks3951.1k
LanguageShellJavaScript
LicenseMITMIT
Last activity23 days agotoday
Topicsskills, codingcoding, orchestration
Curated connections27

autoresearch — the curator's take

The compounding-gains loop, packaged: pick a metric a machine can check, let the agent mutate-verify-keep against it for hours, and small wins stack — the Karpathy recipe without writing the harness yourself. Works across three agent CLIs. The discipline it demands is the catch: without a truly mechanical metric the loop optimizes noise, and unattended iteration burns real tokens — set budgets before you set it loose.

loop-engineering — the curator's take

The methodology layer for autonomous coding agents — read it once you've stopped hand-prompting and want to design the loop that prompts the agent instead: scheduled triage, worktree isolation, maker/checker sub-agents, MCP connectors, human gates, and a phased L1-report → L3-unattended rollout. Ships 7 clone-and-run patterns, tool-agnostic starters (Grok/Claude Code/Codex/OpenCode), and CLIs that score loop-readiness (loop-audit), estimate token spend (loop-cost) and scaffold state/budget (loop-init). It's a patterns + tooling reference, NOT a runtime — it won't execute or host your agents, so bring your own harness (an alook-style AI-company layer or a squid-style pipeline). Overkill if you just want one-shot interactive help; heed its own caveat that unattended loops make unattended mistakes, so verification stays on you.