GitHub Copilot Coding Agent Major Updates and Figma-Codex MCP Integration
Contents
On February 25-26, 2026, GitHub and OpenAI/Figma each announced updates related to coding agents. Five features were added to the GitHub Copilot coding agent, Figma and OpenAI Codex were integrated via MCP, and Copilot CLI reached GA (general availability). A week earlier, Anthropic had announced Claude Code’s Figma integration, accelerating the trend of coding agents being deeply embedded into design tools and CI/CD pipelines.
Five New Features for GitHub Copilot Coding Agent
Model Selector
You can now switch AI models based on task complexity, selecting from a picker UI on the coding agent screen.
Picker options:
- Auto (default)
- Claude Sonnet 4.5 / Opus 4.5 / Opus 4.6
- GPT-5.1-Codex-Max / GPT-5.2-Codex
On the backend, additional models are supported including Gemini 3 series, Grok Code Fast 1, and Claude Haiku 4.5. Each model has a premium request multiplier — GPT-4.1/GPT-4o/GPT-5 mini consume zero, Claude Haiku 4.5 and Gemini 3 Flash are 0.33x, Claude Opus series are 3x, and Claude Opus 4.6 fast mode is 30x.
Auto mode logic currently uses “availability-based model selection” rather than task-based switching. Task requirements will be factored in eventually. Models restricted by admin policies or those with multipliers above 1x are excluded from Auto selection.
Premium request quotas: Free 50/month, Pro ($10/month) 300, Pro+ ($39/month) 1,500. Overages at $0.04/request.
Self-Review
Before the coding agent opens a PR, it runs its own code through Copilot code review (the same one used for PR reviews). It receives feedback, iterates on fixes, then creates the PR.
How code review works internally:
- LLM detection + tool calls: collects project-wide context (code, directory structure, references)
- Combined with deterministic static analysis tools like ESLint and CodeQL
- Detects bugs, performance issues, and overly complex logic
GitHub’s blog featured an example where the agent detected excessively complex string concatenation, simplified the implementation, and then submitted the PR. Review criteria and tone can be customized via copilot-instructions.md.
Security Scanning (Free When Using Coding Agent)
Three scans run automatically before the coding agent opens a PR:
- CodeQL Analysis: detects potential security vulnerabilities
- Dependency Checking: cross-references newly introduced dependencies against the GitHub Advisory Database
- Secret Scanning: detects API keys, tokens, and other secrets
These are paid features of GitHub Advanced Security. Normally, Code Security costs $30/month per active committer and Secret Protection costs $19/month per active committer. For PRs created by the coding agent, they’re provided free — no Advanced Security license or additional setup required.
When issues are found, the coding agent automatically attempts fixes and documents the scan results and remediation in the PR summary. It’s designed to automatically run security reviews on AI-generated code — GitHub’s answer to the concern that “AI-written code is hard to verify.”
Custom Agents
Place files with the .agent.md extension in the .github/agents/ directory to define team-specific agents. The format is Markdown + YAML frontmatter:
---
name: test-specialist
description: "Focuses on test coverage and best practices"
target: vscode # or github-copilot, or both
tools: ["read", "edit", "search"] # ["*"] for all
user-invocable: true
mcp-servers:
custom-mcp:
type: 'local'
command: 'some-command'
env:
API_KEY: ${{ secrets.COPILOT_MCP_API_KEY }}
---
You are a testing specialist focused on improving code quality...Like CI config files, they work seamlessly with git management, reviews, and versioning. Available tools have aliases: execute (shell/Bash), read (file reading), edit (file editing), search (Grep/Glob), agent (call other agents), and web (WebSearch/WebFetch).
Built-in MCP servers include github/* (read-only GitHub tools) and playwright/* (localhost-only browser automation).
Placement has priority ordering: Enterprise .github repo < Organization .github repo < repository’s .github/agents/ < Copilot CLI’s ~/.copilot/agents. Markdown body is limited to 30,000 characters.
CLI Integration and Copilot CLI GA
From the September 2025 public preview, Copilot CLI reached general availability on February 25, 2026. This is a standalone CLI tool separate from GitHub CLI’s gh copilot.
The & key cloud delegation mechanism: prefix a prompt with & to delegate the task to the cloud-based Copilot coding agent. The terminal is freed immediately for other work. /resume seamlessly switches between local and remote sessions. Branch information, logs, and conversation context carry over completely.
Built-in specialist agents:
- Explore: fast codebase analysis
- Task: build and test execution
- Code Review: change review
- Plan: implementation planning
Other features:
- Plan Mode (Shift+Tab) for structured analysis and planning before implementation
- Autopilot Mode for tool execution without approval
/difffor syntax-highlighted change review- Esc-Esc to roll back to previous file snapshots
- Auto-compaction for automatic history compression at 95% context window usage
- Repository memory for remembering codebase conventions across sessions
Models available in CLI: Claude Opus 4.6, Claude Sonnet 4.6, GPT-5.3-Codex, Gemini 3 Pro, and Claude Haiku 4.5 for lightweight tasks.
Figma and OpenAI Codex MCP Integration
What’s Now Possible
Through the official Figma MCP server, Codex can now directly access Figma’s Design, Make, and FigJam files. Announced February 26, 2026.
Key tools exposed by the MCP server:
| Tool | Function |
|---|---|
get_design_context | Retrieves style and layout info for selected layers (defaults to React + Tailwind) |
get_variable_defs | Retrieves variables and styles (colors, spacing, typography) |
get_screenshot | Captures screenshots of selected elements |
get_code_connect_map | Retrieves mapping between Figma node IDs and codebase components |
generate_figma_design | Generates Figma frames from live UI (remote MCP only) |
get_figjam | Converts FigJam diagrams to XML + screenshots |
generate_diagram | Generates FigJam diagrams from Mermaid syntax |
Default framework is React + Tailwind, with support for Vue, HTML+CSS, iOS, and custom components via Code Connect. Rate limits are 6/month for the Starter plan, with per-minute limits for Dev/Full seats.
Comparison with Claude Code Figma Integration
One week before Codex’s announcement, on February 17, Anthropic announced Claude Code’s Figma integration. Called “Code to Canvas,” it captures working UI built with Claude Code from the browser and converts it into editable Figma frames. Not screenshots — text, buttons, and auto-layout structures are imported as editable Figma components.
| Aspect | Claude Code + Figma | Codex + Figma |
|---|---|---|
| Announced | Feb 17 | Feb 26 |
generate_figma_design | Supported | Supported (reports of not being exposed in CLI) |
| Desktop MCP | Local connection | Remote MCP supported |
| Bidirectional flow | Figma → Code, Code → Figma | Bidirectional as well |
Current Limitations
The Figma MCP server exposes visual data like frames, layers, colors, and positions, but doesn’t convey “what it is.” Existing design system prop structures, variant options, and state definitions aren’t directly readable — the coding agent has to infer them. Code Connect can mitigate this by mapping actual codebase components, but it’s limited to Organization/Enterprise plans.
Desktop MCP requires the Figma Desktop App (no browser version, no WSL support). There’s a reported issue where generate_figma_design doesn’t appear in Codex CLI’s tool catalog, with a similar issue on the Claude Code side.
Business logic, event handlers, and state management can’t be carried into Figma layers, requiring manual re-translation when bringing things back to code. A UXPin analysis article noted “Making translation faster is not the same as eliminating it” — effective for greenfield projects, but friction remains for teams with established design systems.
Multi-Agent Comparison with Competitors
| Aspect | GitHub Copilot Coding Agent | Google Antigravity | Cursor |
|---|---|---|---|
| Model selection | 20+ models across 4 vendors (OpenAI/Anthropic/Google/xAI) | Gemini 3 series/Claude Sonnet & Opus/GPT-OSS | Multiple models supported |
| Multi-agent | Custom agents (.agent.md), limited parallel execution | Manager View for parallel agent orchestration | Async parallel sub-agents, sub-agents can spawn further sub-agents |
| Cloud execution | Yes (& for cloud delegation, /resume for handoff) | No (local IDE) | Cloud Agents (independent VM execution, PR generation, reviews with screenshots) |
| Autonomy | Creates PR after self-review + security scanning | Agent-first. Autonomous multi-file operations and command execution | Long-Running Agents complete plan → execution without human intervention |
| Security scanning | Built-in CodeQL/Secret Scanning/Dependency — free | None | None |
| Pricing | $10-39/month + $0.04/request overage | Free (public preview) | $20/month (Pro) |
| Maturity | GA. Enterprise-ready | Public preview. No enterprise auth | GA. Sub-agents added Feb 2026 |
Copilot’s strengths are GitHub integration and free security scanning. Antigravity has the most advanced multi-agent orchestration but lacks enterprise auth and pricing structure. Cursor is catching up with async sub-agent execution and Cloud Agents.