athena-oracle/transcripts/ep_build_agent_os.clean.txt

# How to Build Agent OS - FREE Course! — AI News Today | Julian Goldie Podcast
# Transcript (cleaned) — ~59 min episode
# Source: https://open.spotify.com/episode/4HWHi9beiAW8JGi0CFVgTV
# Resolved RSS: https://feeds.transistor.fm/ai-news-today-julian-goldie-podcast
# MP3: https://media.transistor.fm/618f447d/afb640ea.mp3
# ASR: faster-whisper tiny/int8/beam1 (CPU), transcribed in 12 × 5-min chunks.
# Proper nouns restored. UNCERTAIN terms flagged inline as [??].

INTRO
This is a full one-hour course on how to build your own agent operating system, based on what's working for me. An "agent OS" is a way to have all your agents inside one dashboard, where you build custom workflows and have everything one click away.

EXAMPLE OF WHAT'S POSSIBLE
- A video agent that creates videos in one click.
- An SEO content pipeline: research keywords, generate content, deploy to the website in one click.
- Hermes Jarvis — a voice-operated AI agent that looks and feels how I want.
- Hermes Oracle — pulls trending news.
- Hermes Astros — a custom workflow that takes latest news from Twitter/X, gives trending topic ideas, suggests titles, then plugs into the video agent to automate video; also feeds NotebookLM to create free infographics.
- A lead-gen + email outreach tool: type the leads you want, it generates a list, writes the email campaign, sends it, pulls in lead details, and manages metrics (validated / reachable / sent / replied). Fully automated inbox — one beautiful dashboard.

He notes you can build all this free using your existing subscriptions (e.g. ChatGPT, Claude, Codex — already paid). He used GPT-5.6 [?? "G55.6"] as the brain and "Fable 5" [?? real product?] to help build.

STEP 1 — BUILD THE DASHBOARD
- He uses ChatGPT to build a "Mission Control" dashboard to operate ChatGPT, Codex, and Hermes agent together. Prompt: "create a mission control dashboard where I can operate ChatGPT, Codex plus Hermes agent. Make it clean, beautiful, interesting to use, make sure it actually works."
- You can attach a screenshot of a style you like for inspiration.
- Two install paths: (a) his paid template (updated every time he codes, shipped as a setup/zip file), or (b) build from scratch inside ChatGPT.

STEP 2 — INSTALL & CONFIGURE
- Run it locally on a separate local port. He recommends local over VPS for security: if you give a cloud agent file-edit + computer control and expose it to the world, it's not secure. VPS is fine if you limit the tools.
- "Fast mode" toggle for faster outputs.
- Two instances: template version + from-scratch version. Avoid port overlap (one on port 3200, one on 1200).
- Let it configure parts that don't need an API.

STEP 3 — MEMORY SYSTEM (Obsidian + Omi)
- "Focus on the system, not the model." Hide/replace tools as the landscape changes (he mentions OpenRouter [?? "OpenCore"] was huge two months ago, now less so; GPT-5.6 mind-blowing, but GPT-5.5 not usable for this).
- Obsidian is his preferred memory system — free, open source, stores everything in markdown files in an organized system. Agents organize it.
- Memory galaxy metaphor: every dot = a memory (person/project/business/tool), all linked. Agents understand what you've worked on → better, more personal responses.
- Omi [?? "omie"]: free open-source screen/mic recorder (MIT license, basic tier free). Exports notes directly into Obsidian. "OME → Obsidian."
- You don't build the memory yourself — tell Claude/agents to create and organize Obsidian files.
- Keep the old agent OS running on its port while building the new one (don't let it overwrite).

KEY PRINCIPLE — SIMPLICITY
- The simpler you make it, the better. You won't use every feature. Hide tools you don't use (e.g. he hasn't used "Kimmy code" [?? Claude Code?] in weeks).
- Build only what's relevant to you.

TESTING & ITERATION
- Test it: "are you working?" — he found the from-scratch build created a nice UI but didn't actually test/reply (blank screen). Template version worked immediately.
- Use ChatGPT (or Codex, or Hermes) to build. ChatGPT is more visual. You can fork conversations — run multiple agents in parallel inside the same project folder (e.g. video agent + SEO agent + lead-gen coded side by side).
- Plug in existing subscriptions to avoid API token cost: Claude CLI, Codex (already subscribed) → build and use the Agent OS for free. "You can build these agent cooperative systems for free."

MEMORY INSTALL INTO AGENT OS
- Tell the agent OS: "add the memory system from Obsidian... install it... make sure ChatGPT and Hermes are synced to my memory system. When I use them, update Obsidian; when I ask questions, learn from the Obsidian vault as context." Everything personalized.
- Three layers added in ~27 minutes: front-end, custom workflows, memory.

IMAGE GENERATOR EXAMPLE
- "Create an image of a cat eating biscuits" → Codex figures out aspect ratio (cinematic) on its own. He prefers letting the frontier model make decisions — it has more attention/time than he does.
- Found a real bug (500 error) during testing — valuable. Renaming conversations keeps track of what you're building.
- Token usage warning: fast mode + parallel building burns tokens fast. He hit the usage limit, used weekly reset. Dial down effort/speed/model (e.g. to "4o" [?? "25.5"]) for basic tasks.

TOKEN MINIMIZATION
- 10 strategies in his "token minimization playbook" — claims up to 95% token savings / 2-3× output.
- Templates save tokens (10-20 min vs hours from scratch).

WHAT WORKFLOWS TO ADD
- Do a time audit. Automate anything repeatable, scheduled, that doesn't need you present.
- His content pipeline: keyword research → create → deploy to site in one click (saves ~30 min/week).
- "Based on your time audit, tell GPT-5.6 what you do and have it build custom workflows."

AGENT ORCHESTRATION
- "Paperclip" [?? agent orchestration tool] — a team of AI agents working together 24-7 (Claude, Hermes, Antigravity, Codex).
- Group chat: message a room, agents chip in and bounce ideas off each other, reading the Obsidian vault as shared memory.
- Pipeline system: drop an idea → human approval section → plan → approve → build → gallery of what was built. Idea-to-implementation fast.
- Gallery + workspace: critical so you can find past builds (otherwise lost among 100 projects).
- Claude "ultra-code" [?? subagent feature]: 14 agents working in parallel, Claude as orchestrator building subagents.
- Voice agent (Apollo): listen/talk/build. Full gallery of sessions across all agents.
- "Make it super visual and fun" — small UI touches (icons) make workflows memorable.

MCP & OPEN SOURCE PLUG-INS
- Embed the actual Hermes dashboard inside the system.
- If you see open-source tools, grab the GitHub link, tell ChatGPT "build this into the Agent OS" → code for free. Example: "OmnLyRoot" [= OpenRouter] for free coding; "HY3" [?? a free OpenRouter model] available free until the 21st, then hide/delete it.
- Custom tools: e.g. his own "Figure" [?? a figure/diagram tool] wrapper storing everything in one system.
- Non-business uses: music generator (ambient), goals tracker, etc.
- Kanban board: drop a task → auto-triage → assigned to agent profiles. Changed how his team (SEO agency) works.

TESTIMONIALS / SELLING
- Jacob: set up in 10 min via template vs 12 hrs/day building manually.
- Sheena: built her own version for lead gen.
- For selling AI services: build custom workflows that solve each client problem (image studio, video agent, social media agent).

BUILDING TIPS (what NOT to do)
- Don't run Claude AND Codex on the same thing simultaneously — they can overwrite/delete each other's work. Pick ONE agent to build/improve the OS (he uses Claude + "Faber 5" [??]).
- Don't put it on a public website (Netlify) — keep it secure/local.
- Use free APIs + local models (Ollama, LM Studio) for time-consuming, low-effort tasks. "HY3" free API + local models = near-free operation.
- Document builds on Goldiebench [his benchmark] — also a lead-gen channel.
- Give Agent OS access to Netlify to auto-document the build journal.

CLOSING / PITCH
- System survives model churn: "if Fable 5 [??] gets taken down, in two minutes you remove the integration and add another." Named conversations make it easy to revisit and edit weeks later.
- Voice agent (Apollo) has computer access — can open Google, edit local files, open browser.
- Skills can be plugged in (e.g. "Caveman" [?? a token-reduction skill] reduces output tokens).
- Always give yourself a delete ability for anything that doesn't work.
- Full setup + community (AI Profit Boardroom) inside the Agent OS section: setup/zip file, full guide, daily video tutorials, weekly coaching calls, local member map. "AIProfit.com" [??].

---
NOTE: This 59-min episode had heavy ASR corruption. Confident restorations: ChatGPT, GPT-5.6, Hermes, Agent OS, Obsidian, Omi, Claude, Codex, Antigravity, OpenRouter, Ollama, Netlify, Apollo, NotebookLM. UNCERTAIN (flagged [?]): "Fable 5"/"Faber 5" (a real product he references?), "OpenCore"→OpenRouter?, "HY3"/"25.5" (model slugs), "Caveman"/"Kimmy code" (tool/skill names), "loop engineer" (concept), "Pony's now" (a project name), "Rue sub-brain" (likely "free sub-models"). Promo URLs ("AIProfit.com", community name) were unintelligible in audio.
Raw verbatim: ep_build_agent_os.raw.txt (same dir). Chunks: chunks/c00.txt..c11.txt.