Agents

A runtime for meta-assistants: an agent that can call other agents. Each agent is system prompt + model + tools with its own session, memory and message history. A main agent decomposes a task and hands sub-tasks to specialized sub-agents — for parallelism, specialization, and a smaller main context.

This area is a self-contained platform — it builds and runs on its own and does not require the workflow or semantic-ui areas.

See it in action

A research-lead agent planning a task, spawning web-researcher sub-agents in parallel, having a verifier check their findings, and writing the result to the workspace. The diagram shows the flow; the recording below shows the same flow running live in the Admin UI.

What's in this section

• Getting started — build, set the environment, run the Admin UI.
• Admin UI — the visual console and its sections.
• CLI client — the terminal chat client and its commands.
• Sub-agents — how an agent calls another agent.
• The research-lead flow — a worked end-to-end example with parallel sub-agents.
• Memory — per-agent memory strategies (auto-compression, summarizing window, …).
• Workspace & collaboration — the shared file area, its three scopes, and how agents exchange work.
• Built-in tools — every tool that ships, and what it needs.
• Creating a tool — add your own tool via the SPI.
• LLM gateway — provider routing, retry and throttling.
• Prompt renderer — template variables in system prompts.
• Persistence — the storage ports and the file adapters.
• How it compares — positioning vs. code-first frameworks.

Configuration & reference

• Environment variables — every variable you can set.
• Initial data — how agents and configs are seeded.
• Agent definition JSON and LLM config JSON — the file formats.
• LLM configuration reference — bundled configs and providers.
• Bundled agents — the agents that ship out of the box.
• Modules — the libraries, servers and clients.