12-Page Dashboard
A VS Code panel with dedicated views for agents, teams, skills, and context packs. Browse live state and edit configs without opening a file manually.
Learn moreEverything you need to build an AI agent team in VSCode
Ideal Agent Architecture
Agent Teams is designed around a clean 4-layer pipeline
User Prompt
Natural language via @router
Single entry point
Write @router + any task in plain language. You can reference a file, describe a vague goal, or name a domain — the pipeline handles the rest. No agent knowledge required.
→
Router
Scores & delegates to best agent
Intent-aware dispatcher
Scores every agent by intent keywords, domain vocabulary, and the active file path. Delegates single-domain tasks via handoff and fans out parallel domains simultaneously — without executing anything itself.
→
Orchestrator
Decomposes task into steps
Planner, not executor
Breaks the delegated task into an ordered sequence of subtasks, assigns each to the right worker, and aggregates results. It never writes code or calls APIs directly — coordination is its only job.
→
Workers
Domain specialists execute tasks
Narrow scope, full depth
Each worker owns exactly one domain — backend, testing, docs, and so on. It uses only the tools and skills defined in its spec, returns a structured result, and escalates if the subtask falls outside its scope.
Other valid patterns
Not every setup needs the full 4-layer flow. These lighter patterns are useful too.
Orchestrator → Worker
Planned handoff for multi-step work
User Prompt
Natural language via @router
Complex request enters the system
This pattern starts when the user asks for something that needs planning, decomposition, or coordination before execution.
→
Orchestrator
Plans and delegates the work
Planner in the middle
The orchestrator decides the sequence, splits the task into steps, and hands execution to the worker best suited for the job.
→
Worker
Executes the delegated subtask
Specialist execution
A worker with narrow scope performs the actual implementation, analysis, or update and returns the result upstream.
Only Worker
Direct execution for narrow tasks
User Prompt
Natural language via @router
Small, direct request
This flow is ideal when the task already maps cleanly to one specialist and does not need orchestration.
→
Worker
Handles it end-to-end
Single specialist path
The worker receives the request directly, executes it within its domain, and returns the final result without extra coordination layers.