Architecture overview

Babel is an agent harness for software work with a structured Prompt OS underneath. This page explains the shape of what exists in source today.

Product modes over a shared kernel

The three daily modes — Chat, Plan, Deep — are different entrypoints that share common machinery:

Chat runs a multi-turn ChatEngine loop with a deliberately minimal prompt stack. No orchestrator, no pipeline stages.
Plan runs a governed plan path and stops at a reviewable plan artifact.
Deep loads the governed pipeline (orchestrator, QA review, executor).
All three share the unified execution kernel — completion decisions and tool mapping — so a completion claim is decided by harness policy rather than model confidence.

Prompt OS

The “Prompt OS” is Babel’s structured instruction stack: a behavioral core, domain architects, skills, and task overlays, routed per task from a declared catalog. Chat loads the smallest useful stack. The stack is inspectable before execution — you can preview what would be selected without a model call (see Quickstart).

Interfaces and boundaries

The TUI, headless CLI, Desktop shell, and MCP server are interfaces over the same engine. The Desktop shell runs the CLI; the MCP server is read-only inspection. There is no separate per-interface runtime.

What is not claimed

This page describes implemented boundaries only. No persistent daemon design, no one-runtime refactor, and no future architecture is implied.
Normative architecture detail lives in the product repository; the pages there outrank this summary.