“I was paying double — first Claude, then the API cost — so I asked Cloud Chat to make Cloud Code work like Hermes. I inserted it into Claude Code, and it wasn’t asking anything anymore. One button, it does everything. It worked four hours without touching it.” — Dirk Ohlmeier
Session context: 2026-06-25_Mastermind — Dirk had been trialing Hermes, an autonomous-agent product, balked at the stacked cost, and discovered he could get the same hands-off behavior by feeding Claude Code the right instructions.
Core Idea
When a polished autonomous-agent product impresses you, the instinct is to buy it. But most of what you’re paying for is not intelligence — it’s a behavioral wrapper around a model you may already have access to. Hermes runs autonomously because its system instructions are, at heart, a loop: keep going, evaluate against the goal, if it’s not done try a different way, repeat until criteria are met. Plus a memory layer and a permission posture that stops asking “should I?” every step. None of that requires Hermes. It requires the instructions Hermes encodes.
This is the model-versus-harness distinction made practical. The language model (Opus) is constant; the harness — Claude AI, Cowork, Claude Code — is the system prompt, the memory rules, and the command vocabulary that shape how the model behaves. Dirk’s “miracle” four-hour run wasn’t a smarter model. It was a harness that had been told to loop, double-check its own work, red-team itself, and not stop to ask permission. He inadvertently reprogrammed his own harness to behave like the product he almost rented.
The move, then: when an agent tool wows you, ask “what behavioral pattern is this, and can I express it as instructions in the harness I already own?” Usually you can. Capture it as a skill (or drop it in your global .claude config behind a trigger word — “iterate,” “go until done”) so you can summon the behavior on demand instead of paying a subscription for it. You don’t need the wrapper. You need its loop.
Practical Application
Next time a third-party agent impresses you, before subscribing, run this: “Here’s what this tool does — it runs autonomously, loops until the task is done, double-checks before reporting, and doesn’t stop to ask permission. Write me a skill / a set of Claude Code instructions that makes you behave the same way.” Test it on a real task. Most of the time you’ll find the behavior was an instruction set, not a product — and now it’s yours, reusable, and free. Reserve actual purchases for tools whose value is in proprietary data or compute you genuinely can’t replicate.
Related Insights
- Insight - AI Sovereignty — Build Interchangeable-Intelligence Harnesses So No Vendor Owns Your Workflow — same sovereignty logic, applied to agent products instead of models.
- Insight - The Harness Architecture — Declare Only What You Need, Import the Rest — the harness is the layer you extend to absorb new behaviors.
- Insight - The Model Underneath Is the Multiplier, Not the Interface — the wrapper isn’t the magic; the model and its instructions are.
- Insight - Write Loops, Not Code — Goal-Seeking Loops Are the New Atomic Unit of AI Work — what you’re replicating is a loop.
- Insight - Every Autonomous Loop Needs a Termination Condition — Design It First, Not Last — when you replicate the loop yourself, you own the responsibility for its stop condition.
Evolution Across Sessions
Builds on Insight - AI Sovereignty — Build Interchangeable-Intelligence Harnesses So No Vendor Owns Your Workflow (2026-06-18), which established that the model is a swappable component. New development: the agent product is also swappable — its autonomy is a behavioral instruction set you can lift into your own harness rather than rent.