Execution Model

One active agent coordinated by a deterministic supervisor.

Tars currently runs one active agent. The supervisor serializes inbound work, preserves the active session, invokes the configured model, and routes tool and text events back to the originating channel.

What the supervisor coordinates

• Discord or terminal messages;
• one active conversation history;
• model streaming and tool calls;
• explicit scheduled tasks;
• memory refresh after fact mutations;
• background maintenance and cleanup.

Scheduled tasks use the same supervised execution path. They are not delegated to independent workers or isolated sessions.

No sub-agent orchestration

Tars does not currently spawn specialist agents, route work among roles, or provide a multi-agent task graph. TARS_INSTANCE_ROLE is operator metadata for a named process; it does not create a specialized agent architecture.

If a workflow needs separation, use explicit tools, narrow skills, separate OS accounts, or an external job system. Do not assume model-written delegation provides a security or concurrency boundary.