VS Code 1.109 is Now a Multi-Agent Platform: Claude, Copilot and Codex Working Together 
In the previous post about Copilot we saw how GitHub Copilot evolved from autocomplete to autonomous agent. But in February 2026 something arrived that completely changes the conversation: VS Code 1.109 became a multi-agent platform.
Now you can run Claude Code, Codex and the native Copilot agent at the same time, within standard VS Code, each with its own dedicated context.
It’s not just a Copilot update. It’s an architectural change in how the world’s most popular IDE understands the relationship between developers and AI agents.
What “multi-agent platform” means in VS Code
The difference from what we had before is fundamental. Before, VS Code had one AI chat at a time — you chose Copilot, or installed the Claude extension, and worked with one. If you wanted to switch, you changed extensions.
In VS Code 1.109, the architecture changed:
VS Code 1.109
├── Copilot Agent (native from GitHub)
│ └── Dedicated context: your conversation with Copilot
├── Claude Code Agent
│ └── Dedicated context: independent, doesn't interfere
└── Codex Agent (OpenAI)
└── Dedicated context: independent
Each agent lives in its own context space. The subtasks they delegate don’t consume the main agent’s context — which before was a serious problem for long conversations and complex projects.
AGENTS.md: the instructions all agents respect
One of the most practical features of this update is support for AGENTS.md — a file in your project root that any agent (Copilot, Claude, Codex) automatically reads to understand your project context.
It’s the equivalent of the CLAUDE.md we already know, but universal:
# AGENTS.md
## Project Stack
- Framework: Next.js 15 with App Router
- Database: PostgreSQL with Prisma ORM
- Styles: Tailwind CSS v4
- Tests: Vitest + Testing Library
## Code Conventions
- Components always in English
- Comments and documentation in Spanish
- Server Components by default; "use client" only when necessary
- Explicit TypeScript types always (no `any`)
## When creating or modifying the DB
- First generate migration with `npx prisma migrate dev`
- Update seed.ts if the change affects initial data
- Never modify migrations already applied in production
## Tests
- Every new component needs minimum integration test
- Unit tests for business logic in /src/lib
You create the file once. Any agent working in that repo reads it. You don’t have to repeat the project context in every conversation or with every tool.
The Copilot Coding Agent: real asynchronous work
VS Code 1.109 also consolidates a feature that was in beta: the Copilot Coding Agent now works in an isolated environment (Codespace) and can work completely disconnected from your current session.
The practical flow:
- You open an issue on GitHub: “Add pagination to the users table”
- You assign the issue to
@copilotdirectly from VS Code (new integrated panel) - You close your laptop. The agent works in a Codespace.
- An hour and a half later, you get the notification on your phone: PR ready.
- You review, approve, merge.
For well-defined tasks — bugs, scoped features, specific refactors — this isn’t a lab prototype. It works in real projects.
The key to making it work well: the issue description. The more specific you are (“the createdAt field should display in DD/MM/YYYY format according to the user’s timezone, currently shows UTC”), the better results you get.
Running multiple agents in parallel: how it works in practice
The most interesting part of VS Code 1.109 is the ability to have active conversations with multiple agents simultaneously, each working on a different part of the project.
Real scenario:
Task: Add notification system to a Next.js app
Multi-agent distribution:
├── Claude Code (1M context): "Analyze the complete architecture
│ and design the end-to-end notification system"
│
├── Copilot Agent: "Implement the NotificationBell component
│ and the useNotifications hook according to the approved design"
│
└── Codex Agent: "Write the integration tests for
the /api/notifications endpoint"
Why not use just one? Because each agent has different strengths. Claude Code with its 1M context is ideal for architectural analysis of complex codebases. Copilot is exceptionally strong in implementation integrated with the GitHub ecosystem. Codex has specific advantages in test generation.
The technical key: each agent has its dedicated context. When the Copilot Agent delegates a subtask to a sub-agent, that sub-agent has its own separate context — it doesn’t contaminate or consume the main conversation’s context. This solves one of the most frustrating problems with previous agentic work.
The practical question: when to use which agent?
With three agents available, the obvious question is when to use each one. A practical guide:
Claude Code is your first choice when:
- The project is large or has high technical debt (the 1M context shines)
- You need architectural analysis or design decisions
- You want sub-agents that actively coordinate (Agent Teams)
- The flow is mainly from the terminal
Copilot Agent is your first choice when:
- The task starts from a GitHub issue (unbeatable native integration)
- You want completely asynchronous work without keeping VS Code open
- Your team already lives in the GitHub Actions / Codespaces ecosystem
- The task is scoped and well-defined
Codex is your first choice when:
- The task is intensive in generating code from scratch
- You want a second opinion for a technical decision
Configuration: how to enable multi-agent in VS Code
# 1. Update VS Code to 1.109+
# Help → Check for Updates
# 2. Install GitHub Copilot extension (if you don't have it)
# Marketplace: search "GitHub Copilot"
# 3. Install Claude extension for VS Code
# Marketplace: search "Claude - Anthropic"
# 4. Create AGENTS.md in your project root
touch AGENTS.md
# Add your project conventions (see example above)
# 5. In the VS Code Chat panel
# Open multiple conversations with different agents
# using the agent selector at the top of the panel
Important tip: Although you can run multiple agents in parallel, consider the cost. Each agent runs with its respective model — having three active sessions simultaneously multiplies the cost. The most efficient approach is to use the right agent for each phase, not all at the same time for the same task.
What does this mean for the future of the IDE?
VS Code betting on being a platform — not an editor with an AI plugin — changes market dynamics. Before, users migrated to Cursor or Windsurf to have better agentic capabilities. Now, standard VS Code offers access to the best agents on the market.
What Cursor and Windsurf still offer that VS Code doesn’t: deeper integration with the editor itself (Windsurf’s Cascade writing to disk before approval, Cursor’s parallel sub-agents with isolated worktrees), and a UX designed specifically for agentic workflow.
But for a developer who doesn’t want to change their editor, VS Code 1.109 is now a genuinely competitive option with access to the best models on the market.
Conclusion
VS Code 1.109 is not a minor update. It’s a repositioning change: GitHub (and Microsoft) betting that the future of development is not choosing one agent, but orchestrating several from the same environment you already know.For the LatAm community, this has an important practical impact: if your company already has GitHub Copilot licenses, you now have access to Claude Code and Codex within the same editor at no additional infrastructure cost.
Have you already updated to VS Code 1.109? Are you experimenting with multiple agents? Tell us how you’re using it — especially if you have specific use cases for projects in the region 