A structured path from intent to outcome
The agent does more than pass a prompt to a model. It has routing and context layers that help select the right tools for the current task.
Intent router
Directs greetings, quick actions, short questions and complex requests to appropriate paths.
Context builder
Collects chat summary, preferences, skills, memories and relevant knowledge passages.
Tool router
Selects core tools and adds relevant specialist tools for Windows, Web, Browser, Vision, Knowledge or Trading.
Execute & assess
Runs calls, applies relevant approval policies, continues tool cycles and may verify some success claims.
Relevant information, at the right depth.
The context builder separates short base context from relevant knowledge and deeper task guidance. It is designed to reduce unnecessary prompt size while keeping useful instructions close to the current request.
- ✓BASE — the active chat summary, user preferences and global/selected skills.
- ✓RELEVANT — selected skills, matching memories, prior conversation hints and knowledge excerpts.
- ✓DEEP — supplementary skill guidance and instructions retrieved when a complex job calls for it.
One agent, purpose-built tool families
Available tools change with the request and execution stage. The tool registry describes each tool’s name, inputs, approval behavior and execution function.
Tool registry & specialists
Core tools get priority, while relevant specialist groups can add the tool families most useful for the task. Tools may be dynamically listed or made available by request as supported.
- Windows and system
- Web and browser
- Knowledge and memory
- Vision and trading
Hybrid orchestration
A primary model can coordinate eligible workers, split a task when enabled, build a roadmap, collect outputs and synthesize results. A fallback to the primary model is available for some failed hybrid paths.
- Eligible configured providers
- Optional task splitting
- Provider health tracking
- Combined outputs, not guaranteed superiority
Action planning
When enabled, ActionPlanner can create proposed steps for some multi-step tasks. It is documented as off by default in the supplied guide.
Success verification
The Verifier compares recent action evidence with certain success claims. It is enabled by default in the guide, but some situations are skipped; it is not an absolute guarantee.