Multi-Agent Orchestration

delegating complex tasks to specialized sub-agents.

Overview

Tars utilizes a Supervisor-Orchestrator model to handle complex, multi-step engineering tasks. While the Supervisor manages the primary conversation and system state, it can delegate specialized technical work to Sub-agents.

How Sub-agents Work

Context Injection: The Supervisor identifies a high-complexity task (e.g., “Build a React dashboard”).
Specialized Persona: It invokes a sub-agent with a dedicated system prompt tailored for that specific domain.
Autonomous Execution: The sub-agent runs in an ephemeral, isolated session. It has access to the same toolset as the Supervisor but focuses entirely on the technical objective.
Handoff: Once the task is complete, the sub-agent’s findings or artifacts are returned to the Supervisor, who then presents them to the user.

Built-in Agents

Scaffolder Agent

The scaffolder is a specialized agent for structural project setup. It is capable of:

Initializing project boilerplates (Next.js, Vite, Node.js).
Configuring complex tsconfig.json or eslint setups.
Creating standardized directory structures for new features.

Creation & Management

Sub-agents are stored as markdown files in ~/.tars/.gemini/agents/. Each file contains the system instructions for that specific persona.

Tars can create new sub-agents at runtime if it identifies a repeating persona that would benefit from specialized isolation.

Benefits

Reduced Hallucinations: By focusing the context on a specific domain, sub-agents are less likely to conflate unrelated information.
Context Hygiene: Sub-agent interactions don’t bloat the main user conversation, keeping the response speed high.
Parallelism: In future updates, Tars will support spawning multiple sub-agents to work on different parts of a project simultaneously.