The most capable agent system is not necessarily the most complex one. For local and embedded deployments, every additional service creates another failure mode, permission boundary and operational dependency. Our working model therefore begins with explicit control rather than orchestration theatre.
A small, observable operating model
The interface is a terminal UI. It keeps the current task, available tools, intermediate outputs and errors in one inspectable place. A single supervisor model coordinates specialised workers instead of allowing autonomous workers to exchange uncontrolled messages.
The supervisor has three responsibilities:
- decide whether a task needs a tool, a model call or human review;
- pass narrowly scoped instructions and structured inputs to a worker;
- validate the result before it becomes an action.
This design is deliberately close to Unix: small components, clear inputs and outputs, and a visible execution path.
Five rules for local agents
- Tools are explicit. Every capability has a name, a schema and a permission boundary.
- State is inspectable. A user can understand why a decision was made without reconstructing hidden conversations.
- Workers are replaceable. A classifier, planner or retriever can be exchanged without redesigning the entire system.
- Actions are reversible where possible. Side effects are logged and require confirmation when they cannot be undone.
- Evaluation comes first. Agents are tested against fixed tasks, tool traces and failure cases before they are trusted with live work.
Files as an event log
For small systems, append-only files are often a better starting point than a database. They are portable, easy to inspect and straightforward to version. A structured event log records prompts, tool calls, outputs and decisions. A database becomes useful when query volume, retention or concurrent access genuinely requires it.
Conclusion
Reducing the architecture is not a reduction in ambition. It creates a system that can be measured, audited and improved. That is the foundation for useful autonomous agents on constrained hardware.