How I Get Fable 5 Level Results with Any Model (Seriously) Using AI Harness Engineering
Fable 5 was incredible, then it got pulled. I didn't lose a single workflow, because the part that did the work wasn't the model. It was the harness.
You don't need another AI tutorial. You need a community full of people who are actually building. Join the AI Captains Academy: https://skool.com/aicaptains
Ready to go deeper? Free 5-day email courses:
→ AI Foundations Sprint: https://foundations.aicaptains.ai
→ Command Line Bootcamp: https://bootcamp.aicaptains.ai
An AI agent is really a model plus a harness. The model is the brain (the intelligence you rent). The harness is the body — the context, tools, checks, and memory that turn any LLM into something that actually ships work instead of just chatting. That body is the part you own.
In this one I break down what an agentic harness is, walk through the agent loop (read → pick a tool → run → check → done?), and show two of mine running live: my content forge (which cooked for ~35 minutes with no babysitting) and a directory harness I build from scratch on camera for sovereigntyatlas.com. You'll see the real rough edges too — I close a running harness by accident, a draft lands in the wrong place, and my voice-reviewer agent catches unsourced claims and forces a rewrite. That's the honest version.
The future belongs to builders who can create their own solutions. Stop depending on others' platforms and start building your own digital empire. Become an AI Captain.
🔗 Resources:
Safe Agentic Workflow Github: https://github.com/bybren-llc/safe-ag...
Opencode: https://opencode.ai/
VSCodium: https://vscodium.com/
Venice AI: https://venice.ai/home
⏱️ TIMESTAMPS:
0:00 - Fable 5 got pulled
0:39 - Why this channel — sovereign AI
1:03 - The directory I built right before Fable vanished
1:50 - What an agentic harness actually is
2:17 - The agent loop: read, pick, run, check, done
3:04 - An agent = a model + a harness
3:58 - Rent the brain, own the harness
4:27 - The two problems with renting the model
6:05 - Build with frontier, execute with open source
7:11 - Your job moved up: you're the director now
7:34 - Stealing a harness: the SAFe Agentic Workflow
8:55 - My content forge running live (~21 min unattended)
11:11 - Anatomy of a harness: the knowledge base
11:48 - The .claude folder — agents and skills
13:08 - Hooks, commands, and CLAUDE.md
14:55 - Context engineering with sub-agents
17:15 - Swap the brain, keep the body
17:32 - The AEO audit rig (a 10-agent pipeline)
18:55 - What the forge produced (and what broke)
21:08 - How to get started: steal a good one
21:55 - Building a harness from scratch on Sovereignty Atlas
26:07 - Running it fresh — the handoff prompt
30:01 - Results: the directory updates itself
31:03 - The voice reviewer catches unsourced claims
33:36 - That's a harness — recap
34:26 - Start by stealing a good one
34:48 - AI Captains Academy + close
The move underneath all of it: build with frontier models (Opus, GPT, Gemini), then execute with cheaper open-source ones. Swap the brain, keep the body. That's a durable asset — and you don't need to be a developer to own one.
For fellow builders who'd rather own the rig than chase the next hyped model.
Say hi on X: https://x.com/jordanurbs
Substack: https://jordanurbs.substack.com/
For the algorithm:
Claude Fable 5
AI Harness Engineering
AI Harness
Harness engineering ai
What is harness in ai