The harness
What sits between a model and real work, and why it should be open.
A language model on its own does nothing. It emits tokens. Everything that turns those tokens into work is the harness, and for the last two years the harness has been where the real engineering went and where almost none of it has been public.
What the harness actually covers
When someone says “an agent did X”, four separate systems were involved and only one of them was a model. Naming them separately is the whole design.
- Routing
- Deciding what the request means and who should handle it. Not the same as answering it. In Vibemind this is the Brain, and it tries to answer deterministically before it asks a model anything.
- Execution
- Actually running the work, with a time limit and a record of what happened. In Vibemind this is OpenFang, a Rust agent runtime. It ships a metered WebAssembly sandbox, though no Vibemind agent runs inside it today — most run as Claude Code subprocesses bounded by a timeout.
- Memory
- Keeping two different things: what the system decided before, and what your material actually is. These are different questions and Vibemind keeps them in different stores on purpose.
- Transport
- Getting the intent in and the answer out, across whichever channel the person happened to be using. Today each channel is received and answered by the same gateway: Telegram by OpenFang, WhatsApp by OpenClaw.
The design bet
The routing layer should be boring. A regex match against a capability registry runs first; anything model-based is the fallback, not the default. This is unfashionable and it is the point — a regex hit is free, instant, and you can read why it fired. A model-based decision costs a call and produces an answer you have to trust rather than verify.
The consequence is that the interesting behaviour of the system is inspectable as configuration rather than emergent from a prompt. You can read the capability registry and know what the system can route. You cannot do that with a prompt.
What open means here
Open means you can read the routing configuration and see which capabilities are enabled, clone the runtime and run it on your own hardware, and check any claim on this website against the code that implements it. It does not mean every component carries the same licence — some are copyleft, one is non-commercial, a few have no licence file yet. See Licensing for the honest breakdown, which is messier than we previously said.