Research: Agent Loops, Codex, and Claude Code

June 8, 2026 · View on GitHub

Date: 2026-06-08

Takeaway

Latest loop hype points at one useful shift: stop hand-prompting every agent step, and start writing durable control flow that prompts agents for you. Practical version is not "agents decide everything." Practical version is a state machine around specs, prompts, verification, review, and backpropagation.

What "Loop" Means Here

A loop is deterministic control flow around probabilistic workers:

input queue: specs, issues, incidents, PRs, failing checks
planner: decides execution shape from explicit rules
worker: Codex or Claude Code agent/subagent
verifier: tests, lint, screenshots, logs, review agents
state store: files, manifests, PR comments, run logs
backpropagation: facts learned from implementation update specs/docs
stop rule: archive accepted work, request decision, or retry bounded failure

Source-Backed Platform Notes

OpenAI Codex:

Skills are reusable workflow packages with SKILL.md, optional scripts/references/assets, and repo discovery through .agents/skills (OpenAI Codex manual, Agent Skills).
Codex subagents can spawn specialized agents in parallel, but only when explicitly asked; custom agents live under .codex/agents/ or ~/.codex/agents/.
Codex hooks run scripts on lifecycle events such as UserPromptSubmit, PostToolUse, SubagentStop, and Stop; project hooks live in .codex/hooks.json or .codex/config.toml.
openai/codex-action@v1 runs Codex in GitHub Actions for review, patches, and repeatable CI tasks.

Claude Code:

Skills create slash-invokable commands and auto-loaded workflows from .claude/skills/<name>/SKILL.md.
/loop reruns prompts on a schedule inside a Claude Code session; .claude/loop.md customizes bare /loop.
Dynamic workflows move orchestration into JavaScript written by Claude and executed by a workflow runtime; use for dozens to hundreds of agents, codebase audits, large migrations, and cross-checked research.
Agent teams are experimental multi-session coordination with peer communication; subagents are lower-overhead single-session workers returning summaries.
Hooks can enforce lifecycle guardrails at tool, prompt, subagent, task, and stop events.

What Tweets Get Right

The useful framing:

Living specs as code beat one-off prompts.
Folder queues are simple, inspectable orchestration.
Backpropagation keeps docs/specs from rotting.
Humans should decide priorities and product meaning.
Agents should execute bounded work, generate code, run checks, and summarize evidence.

Failure Modes

Infinite loops without hard stop rules.
Hidden product decisions in auto-generated prompts.
Huge fan-out with no cost budget or review gate.
Specs treated as write-only prompts instead of durable contracts.
Review agents summarizing instead of checking diffs and tests.
Backpropagation changing intent instead of implementation facts.

Loop Factory Mapping

Need	Repo Primitive
Living spec queue	`factory/specs/inbox`, `active`, `archive`
Repeatable prompt generation	`loop-factory dispatch`, `prompt`, `review`, `backprop`
Codex-native workflow	`.agents/skills`, `.codex/agents`, `.codex/hooks.json`
Claude-native workflow	`.claude/skills`, `.claude/agents`, `.claude/loop.md`
CI review	`.github/workflows/codex-review.yml`
Backpropagation	`loop-factory backprop` prompt from git diff
Built-in loop catalog	`loop-factory loops list/show/prompt`

Built-In Loop Catalog Pattern

Catalog-style loop libraries make loops discoverable by name, category, trigger mode, supported agents, check command, and exit condition. Loop Factory adopts that shape for original repo-native loops, while keeping wording and behavior tied to its own spec queue.

Current built-ins cover:

spec grilling before dispatch
spec dispatch
active spec verification
independent review gate
spec/docs backpropagation
cleanup pass
docs sync
failing test triage
build repair
CI branch watch
PR health watch
dependency audit planning