Agent orchestration is a pattern where one AI agent coordinates multiple specialized sub-agents (planner, coder, designer) to complete complex tasks. Using GitHub Copilot CLI and VS Code custom agents, you can create a lightweight orchestration framework where each sub-agent uses a different model optimized for its role—Claude Sonnet 4.5 for orchestration, GPT-52 for planning and coding, and Gemini 3 Pro for design.
When you use AI tools today, you might have multiple agents running—a local agent, a background agent, maybe a cloud agent—but you are the orchestrator, manually managing them all. Agent orchestration flips this: you have one AI agent that automatically calls and coordinates other agents.
This is possible with recent tooling in GitHub Copilot CLI and VS Code custom agents, which allow sub-agents to be called by a main agent, each potentially using different models.
In Copilot CLI, you can simply tell the agent to consult multiple models. The CLI will spin up parallel reviews across models automatically. Most people don't realize you can have one model call others within a single Copilot session.
Burke demonstrates an "ultralight orchestration framework" with four custom agents in VS Code:
The orchestrator's only tools are the agent tool (to call sub-agents) and memory (new Copilot feature). The prompt is simple: "You're a project orchestrator. Break down complex requests into tasks and delegate to specialist sub-agents. You never implement anything yourself."
Critical rule: Tell the orchestrator not to give sub-agents detailed instructions on exactly what to do—otherwise it micromanages and the sub-agents can't use their expertise.
The Planner uses GPT-52 for creating high-level plans without writing code. The Coder uses GPT-52 Codex with Context7 MCP for documentation access, and is told to "question everything" from the orchestrator. The Designer uses Gemini 3 Pro with full creative autonomy for UI/UX work.
The most fascinating result: 2,771 lines of code generated using only 10.8K of context window. Sub-agents don't pollute the main context because each has its own isolated context window that disappears when done.
Burke identifies that the planner should save plans to a document that gets fully passed to the coder, and the orchestrator should parallelize coding work—spinning up 5 coder agents on discrete chunks instead of one doing everything sequentially.
"In this scenario, you are the orchestrator. You are running all of these different agents. But in an agent orchestration scenario, you actually have one agent that calls others."
"We want the agency of Sonnet and we want the coding chops of Codex."
"These models think they know everything and so you have to really go out of your way to make sure that they don't tell sub-agents exactly what to do."
"Look at this—2,771 lines of code and we've only used 10.8K of the context window. How is that possible? That's the magic of sub-agents."
| Time | Topic |
|---|---|
| 0:00 | Introduction |
| 0:36 | What is orchestration? |
| 3:38 | Orchestration with custom agents |
| 5:38 | Orchestrator Agent |
| 7:23 | Planning Agent |
| 7:57 | Coder Agent |
| 8:58 | Designer Agent |
| 10:10 | Orchestration in action |
| 15:09 | Improving the orchestration |