What Are Loops — Agents, Orchestrated · EP 01

A ~58-second LinkedIn motion graphic on a pattern every agentic coding tool has quietly converged on: the loop — an agent that runs itself (generate → run → check → repeat). Claude Code, OpenAI Codex, and Google Antigravity all shipped the same /goal-style command, independently, within months. The take: autonomy is table stakes; verification is the moat — the loop you can audit is the only one that ships.

The meta part (why this repo exists): I didn't hand-edit this video. I pointed a single Claude Code /goal at a spec — VIDEO_SPEC.md + CAPTION_SCRIPT.md — and it built the whole thing, then verified itself frame-by-frame (rendering stills and checking the right caption was actually on screen) before calling the job done. A verification loop, to make a video about verification loops. This repo is the spec, the verification approach, and a deep-dive on the concept — not the rendering code.

📺 The video lives on LinkedIn → https://www.linkedin.com/in/zishanalikhan/

---

Start here

• LOOPS_DEEP_DIVE.md — the long version of the video: what a loop is, the four components every loop is built from, and how each company (Claude Code · Codex · Antigravity · Cursor) does it right now, with a side-by-side table. ← read this
• VIDEO_SPEC.md — the exact spec I handed the /goal: objective, design system, storyboard, and the Definition of Done.
• CAPTION_SCRIPT.md — every on-screen word + the 4-layer verification loop that decides "done."

What's inside

Path	What it is
LOOPS_DEEP_DIVE.md	Field guide: the anatomy of a loop + how each tool implements it.
VIDEO_SPEC.md	The build spec — and the /goal that runs it.
CAPTION_SCRIPT.md	Locked on-screen copy, timing, and the 4-layer verification.
NARRATION.md · teleprompter.html	Voiceover script + a synced teleprompter for recording it on time.
Screenshots/	The real product captures — the only real third-party UI in the video (my own screenshots).
out/loops-vertical.mp4 · out/map-2x2.png	The rendered video and the 2×2 map.

The Remotion implementation that renders all this is kept private — this repo is the spec + the deep-dive + the result.

How "done" is decided (the 4-layer loop)

The build isn't finished when an .mp4 exists — only when all four layers in CAPTION_SCRIPT.md pass, in order:

0. Build & static — compiles clean, both compositions register, every asset resolves.
1. Frame stills — render a still per scene and confirm the correct locked caption is on screen, legible, not clipped (evidence, not vibes).
2. Content & accuracy — one shared GOAL_TEXT typed identically in all three tools, every locked string verbatim, the 7 map pills correct, only my own imagery.
3. Full render — both .mp4s render error-free, ~58 s, audio synced.

Guardrails (non-negotiable)

Only my own screenshots and my own headshot appear as real imagery. The /goal command is typed into clean input fields laid exactly over each tool's real input box. No OpenAI / Anthropic / Google copyrighted figures, docs, or marketing assets are embedded. The conceptual graphics — the loop primer, the "anatomy of a goal," the prompt-vs-goal contrast, the two-axes map, the verify-recap — are original, inspired by OpenAI's Goals cookbook, never copied from it.

Series

Part of Agents, Orchestrated — a weekly series on what makes agentic systems actually ship (not chatbot demos). Each episode pairs a broad-reach topic with a deep take on the control plane: verification, human-in-the-loop, governance, orchestration. Autonomy is table stakes; the moat is everything around the agent.