An agent does not need a permanently loaded model.
A logical agent preserves its identity, role and history between tasks. Physical reasoning capacity can be shared across eligible jobs, with model residency managed separately.
This separation is central to Compacitas: organize a company around responsibilities, then allocate execution capacity around the work that is actually ready to run.
Qualify first. Optimize within the boundary.
The local Multiplex core checks identity, permitted capabilities, model allowlists and resource budgets before reserving work. Its scheduler supports priorities, aging, deadlines, admission and backpressure.
Eligible candidates can then be ranked using scoped quality measurements, observed service time, configured cost, residency and failure penalties. This is evidence-informed routing; it does not imply global optimality or an automatic judgment that every answer is correct.
- Persistent identities and a durable work queue
- Mission budgets for time, tokens, configured cost and attempts
- Model reuse and idle-adapter eviction
- Versioned claims, recovery and execution records
Explain the placement decision.
A useful execution record connects the requesting agent, mission, selected substrate and result. Routing explanations should preserve why the selected option was eligible and what measurements informed the decision.
Quality, compatibility and execution are separate claims. A completed model call alone does not prove task quality.
A clear implementation boundary.
The inspected local execution core implements scheduling, budgeting, recovery and routing mechanisms. Its local SQLite scheduler is not a distributed scheduler or a hardware resource clamp.
The wider node-to-cluster product architecture depends on its deployment configuration. This product tour does not announce a production GPU benchmark, automatic model creation or a qualified distributed deployment.