Skip to main content

Architecture Overview

The operator model is the center of this SDK.

The operator does not define prompts, browser steps, or hidden workflow logic. The operator defines:

  • a goal
  • a capability
  • optional structured context
  • optional runtime
  • optional verify logic

Why this exists​

This split keeps the system readable:

  • task meaning lives in the capability
  • execution input lives in context
  • execution resources live in runtime
  • execution behavior lives in the engine
  • acceptance logic lives in verification

How it works​

OPERATOR
  │
  │ task + context + verify
  ▼
Capability-selected execution
  │
  ▼
Engine
  │
  │ uses runtime
  ▼
Runtime (Browser / APIs / Test Runtime)
  │
  ▼
Environment
  │
  ▼
Verification
  │
  ▼
Result + Trace

Current​

Current

Today, operator-runtime supports:

  • a typed operator API
  • capability registries
  • capability default verification
  • per-run verification
  • execution traces
  • a browser-backed Playwright example
  • a browser-backed integration test

Planned​

Planned

The docs refer to future engines and runtimes so the architecture stays extensible. A full LLM-backed runtime is not implemented in the repo today.

Notes​

  • A capability is not a browser method.
  • Context is not prompt text.
  • Runtime is not task meaning.
  • Engine is not the same thing as runtime.