Local models
Local inference, for certain tasks or for the whole harness.
respawn is an agentic harness that primarily runs on frontier model APIs. Local inference is a setup the same harness can use: certain tasks run on local models, or the entire harness runs on local models. The business context stays on the server the business owns.
Three ways the same harness can run
Frontier model APIs
The usual setup. The harness runs on hardware the business owns and uses frontier model APIs for the model work. Your data, workflows, tools, and secrets stay on hardware you own.
Local inference for certain tasks
Local inference is set up for certain tasks. Those tasks run on local models, and the rest of the work uses frontier model APIs. The business context stays on the server the business owns.
Local inference for the entire harness
The entire harness runs on local models. The harness, the models, and the business context are all on hardware the business owns.
In every setup
What stays on the server
The setup changes where the model work happens. It does not change who runs the harness or where the business keeps what it knows.
- Data
- Workflows
- Tools
- Secrets
Your data, workflows, tools, and secrets stay on hardware you own. The business context stays on the server the business owns.
Sub-agents
Sub-agents on frontier models or on local models
An agent can spawn sub-agents that use other frontier models, and sub-agents that use local models. On the usual path, a sub-agent uses another frontier model. Where local inference is set up, a sub-agent can use a local model.
The harness that spawned them is the one on the business’s server, and the business context stays there with it.
Choosing
A real second path
Frontier model APIs are the usual path. Local inference is a real second path: an office can set up local models for certain tasks, or run the entire harness on them.
Either way it is the same harness. The procedures, the tools, and what it has learned stay where they are.
Ownership
Swap the brain anytime.
respawn is reasoning-engine agnostic — swap the model when a better one appears. The harness, the workflows, and everything it has learned can be brought back without starting over.
That holds for frontier models and for local models. The business context stays on the server the business owns.
Demonstration
Run a job on a frontier model, then with local inference.
The demo has the same three choices. Run a job each way and compare which steps used which model.