Codex (OpenAI)
First met in aicp-c3-basecamp-03 · Coding agents & editors
Codex (OpenAI)First met in aicp-c3-basecamp-03 · Coding agents & editors
OpenAI's coding agent, run at GPT-5.5 'extra high' — the course's pick when back-end correctness matters more than looks.
You need generated config or back-end code to be right first time — it produced the only n8n JSON with native Firecrawl nodes when Claude missed them.
Reasoning effort is a dial you turn up — 'extra high' was the setting that won both BC3 and BC4 comparisons
Live preview while building
Strong at structured output (JSON, workflow definitions) where a near-miss is a broken artifact
Generates and repairs structured config — n8n workflow JSON, schemas
Builds full projects with a live preview and pushes them to GitHub
Runs an error-fix loop: paste the failure, get the corrected artifact
Doesn't design well — 'strong back end, average front end' in the bake-off
Doesn't replace review of generated automation; a wrong node is a silently broken workflow
The generate-the-config seat: workflow JSON, schemas, back-end logic. Pair with a vibe-coding platform for anything user-facing.
Coding agents & editors
No current need for a second coding agent alongside the Claude stack.
Worth remembering as the second opinion when a generated artifact has to be exactly right — the BC3 Firecrawl-node miss is the cautionary case for trusting one agent.
current
claude-code: the head-to-head the course keeps running; Codex won on n8n JSON
n8n: 'use AI to build AI' — generate the workflow JSON, then import it
| Session | Coverage | Moment | What happened |
|---|---|---|---|
| Basecamp 3: Introduction to n8n — I | demonstrated | 2:30:17 | GPT-5.5 'extra high' produced the better JSON with native Firecrawl nodes; also the error-fix loop; trainer mixes Codex with n8n to build workflows |
| Basecamp 4: Introduction to n8n — II | demonstrated | 2:21:38 | GPT-5.5 'extra high' with live preview; strong back end, average front end; repo pushed for learners; trainer's YC early access; also his n8n-local-install helper; has 'pets' feature |
| Session 3: Building Elegant Landing Pages | demonstrated | 0:09:38 | The session's workbench end-to-end: projects, GPT-5.5 reasoning tiers, files/browser/terminal panel, git auto-tracking, skill installs, automations, plugins; his team migrating to it from Claude Code |
| Session 4: How to Build Micro Prototypes | demonstrated | 1:37:55 | Open-existing-folder projects, 'enable git and push', run-the-server local preview (localhost:8080), commit/rollback demo, prompt queueing, multiple windows (4 apps at once) |
| Session 7: The Outreach Engine | demonstrated | 0:45:35 | Built both skills live ('GPT 5.5 extra high'); chosen because it writes files straight to his desktop — also organizes his desktop weekly |
| Session 8: The Conversion Playbook | demonstrated | 1:16:34 | Rebuilt his live /consulting page mid-session, created the /coaching short redirect, installed the tracking pixel (head injection), built the ICP-finder skill — all pushed live via GitHub |
| Session 14: Claude Code — Advanced Workflows | demonstrated | 0:09:22 | His own studio shown: one autonomous 1d 1h 17m run on 'GPT 5.6 Sol' (as heard); YouTube API auto-publish configured |
| Session 15: Loop, Harness & Graph Engineering (posted as 'Securing Vibe-Coded Apps') | demonstrated | 2:49:59 | /goal available; /loop not (as stated); starring as the different-model JUDGE in his graph — agent-evals bias argument |
| Session 18: Nano Banana + Veo 3.1 Multi-Shot Video — UGC Genie, an AI-UGC Product Built Live | demonstrated | GPT-5.6 Sol light for speed; annotate tool, in-app browser (Atlas absorbed), side chat, agent's own cursor doing browser checks | |
| Session 19: Content Strategy Pipeline — the YouTube Growth Stack, Built by an Agent Team of Chats | demonstrated | Orchestrator + sub-chats + worktrees + /side chat + status mini-sites; running on API credits, not the app plan | |
| Session 20: Complete LinkedIn Agency Automation — Carousels + Newsletter from One Content System | demonstrated | GPT-5.6 extra-high for planning, Sol light for speed; multi-chat delegation incl. the GitHub force-push chat | |
| Session 21: Advanced Data Extraction from PDFs — Parsing, OCR, and the Token-Free Product | demonstrated | Build agent: brainstorming skill, writing-plan skill, spec docs, thin adapter, three parallel chats by session's end | |
| Session 22: Automate Admin Tasks — Advanced RAG (Enterprise 'Internal Perplexity' Built Live) | demonstrated | Primary build agent; queueing, steering, side chats, Connections-to-VPS all demonstrated; worktrees puzzled over live | |
| Session 23: Automating AEO/GEO — Audits, the File Layer, and Evidence-Led AI Search | demonstrated | The working agent throughout — skills, audits, fixes, the tool build; GPT-5.6 tier switching live | |
| Session 24: Building a Marketing OS — Sector-Agnostic Command Center on an Open-Source Stack | demonstrated | The recovering agent: reads the handover, one-shots the mini-site + Velora build, drives PostHog/dashboard integration | |
| AI Sprint: The Hermes Agent — Day 1 (Always-On Agents, the Employee-on-Day-One System, and the Self-Improving Ghostwriter) | demonstrated | The harness: installs Hermes, wires Ollama, researches file conventions, fixes keys, debugs, and finishes the image skill when Hermes stalls | |
| Session 9: Introduction to Cursor & its principles | mentioned | 0:37:46 | This month's main; local-system access (no sandbox); can open its own browser past CAPTCHAs |
| Session 11: Overview of Claude Code & Key Concepts (Installs, Modes, Context Files, Skills, Hooks & MCPs) | mentioned | 2:09:02 | Named competitor (OpenAI); can't nest inside Claude Code, though GPT models can be used within it |
| AI Sprint: The Hermes Agent — Day 2 (Loop Engineering: the Scheduling Loop, the Delegation Loop, and the Skill Auditor) | mentioned | Sub-agent delegation target in Hermes' skill pack; 'Codex doesn't have sub-agents, I think' aside | |
| AI Sprint: The Open Source Playbook — Day 2 (Weights 101, OpenCode, and the Subsidy Math) | mentioned | The other 50%; also 'does parallel sub-agents very well' | |
| AI Sprint: AI Evals & Reliability — Day 2 (Testing Systems That Never Answer Twice) | mentioned | The other tool behind the certificate generator |