From "Harness" to "Compass".
Half a year ago, when models were less capable, the harness made sense — restrain the output, prevent harm. But Opus 4.5-class models can already match a well-briefed senior engineer when they are given proper direction. Tying them down with a harness only suppresses what they can do. What they need now is a solid map and a clear compass — so the latent ability of the LLM can actually come out.
You don't put a harness on a senior engineer; you brief them and trust them. AI deserves the same now. The Compass is not a fence — it is a north star the agent can refer to whenever it has to make a local decision, and a map of the terrain it is operating in. Give it both, and the agent stops asking permission for every step and starts shipping.
Micromanaging a capable model is not safety — it is throughput tax. Every extra rule you bolt on is an instruction the agent has to re-read on every step, an opportunity for the answer it would have given for free to be twisted into something worse. The harness era is over because the cost of restraint now outweighs the risk it was meant to control.
Harness
- Restrain output to safe forms
- Block, gate, sandbox
- Optimized for early, unreliable models
Compass
- Give direction; trust judgement
- Brief, then let move freely
- Optimized for capable, fast-improving models
The Intent Compass template
A single artifact at the root of every domain. Short. Curated. Read by every agent on every run.
# Intent Compass — <domain> ## Direction What this domain is trying to achieve, in one paragraph. ## Boundaries - What is in scope - What is out of scope ## Trade-offs we accept - We prefer X over Y when … ## Open questions - …
Why this matters — our conviction
If you tie down an Opus 4.5-class model with a harness, you will never see what it is actually capable of. You will see a smaller, slower, more apologetic version of it — and you will conclude that AI coding still isn't there yet. That conclusion is wrong. It is the harness talking, not the model.
Hand the same model a real map and a clear compass — what the system is for, what the boundaries are, what trade-offs we have already made — and the latent ability inside the LLM finally has somewhere to go. It moves on its own. It makes the local calls a senior engineer would have made. It comes back with work you actually want to merge. That is what we are after.
"The harness was right when models needed to be stopped. The map and the compass are right when models can be trusted to walk." Stop restraining capable AI. Brief it, point it, and let it draw the line itself.
Other vocabularies we are testing
"Compass" is the term we are leading with, but the underlying shift — from restraint to direction-and-trust — can be named in several ways. We are gathering reactions to these alternatives; if one of them lands harder for your team, please tell us.
- Navigation — Share the destination and the current position; let the agent pick the route.
- Compass — Show the direction that matters; leave the moment-to-moment navigation to the agent.
- Mission Charter — Hand over a mission and a definition of done; how to execute is the agent's call.
- Empowerment — Give the information and the authority needed to act independently.
- Trust Delegation — Delegate on the basis of trust and stay focused on the outcome rather than the steps.
Internally we still default to "Compass" — it is short, it is concrete, and it pairs cleanly with "map" (the structured intent tree it points at). But the choice of word matters less than the shift behind it: stop restraining capable AI; start trusting it with direction.