Stacks/Coding agent
The coding harness behind GitHub Copilot in VS Code
The VS Code team's account of Copilot's editor-side coding harness treats the model as the engine and the harness as the car. The post walks through context assembly, tool exposure (including MCP and custom agents), the tool-calling loop, and a dedicated eval pipeline (`vscode-evals` and `github/evald`) that benchmarks harness changes on every relevant PR.
- Use case
- The agentic coding experience inside Visual Studio Code: assembling context, exposing and executing tools, and running the agent loop that turns model output into editor actions
- Category
- Coding agent
- Team size
- not stated
- Orchestrator / framework
- A tool-calling agent loop (think, act, observe, think again). The harness builds the prompt (system instructions, user query, workspace structure, conversation history, tool results, custom instructions, and memory from earlier sessions), sends it to the model, executes any tool calls (read_file, replace_string_in_file or apply_patch, run_in_terminal, semantic_search, and more), and repeats until the model returns no tool calls. The available tool set can change per request, including MCP servers, extensions, user toggles, and custom agents defined in `.agent.md`.
- Memory
- The context assembly step includes memory from earlier sessions, along with conversation history and custom instructions.
- Evals
- A CI path for harness changes: a successful build publishes a versioned eval agent to a `vscode-evals` npm feed, which opens a model-evaluation issue on `github/evald`. evald runs the benchmark and reports analysis back to the original VS Code PR via an Azure Logic App (comment URLs only; the analysis body stays private).
- Observability
- A Chat Debug View in VS Code lets developers inspect prompts, tool calls, and results behind an agent run. A Tools UI shows which tools are available for a request.
- Monthly cost
- not stated
- Models
- not stated
- Observability tools
- Other tools
- MCPCustom agents (`.agent.md`)github/evaldAzure Logic App
- Source
- https://code.visualstudio.com/blogs/2026/05/15/agent-harnesses-github-copilot-vscode
Julia Kasper, Megan Rogge, and Aaron Munger · May 14, 2026
Related
More like this
Coding agent
Date not stated
Stripe Minions
Unattended end-to-end coding agents at Stripe. Over 1,300 merged pull requests each week are completely minion-produced and human-reviewed, with no human-written code in those PRs
Coding agent
Sep 28, 2026
Replit Agent (model-directed harness)
Replit's production coding agent, whose core loop decides how hard to think, when to hand work off, and which subagent (and model) should do it
Coding agent
Apr 22, 2026
Cognition Devin cloud agents
Cloud software-engineering agents that work unattended across the async gaps of real development — opening a PR, waiting on CI, responding to review, and pushing follow-up commits