Log in Sign up
News XREAL's Aura Puts Android XR and Gemini Into 95 Grams of Glasses for $1,099 Business AI Hyperscalers Quintupled Flash Prices, and Apple Is Paying $25 Billion for NAND It Once Bought for $6 Billion Coding Assistants Microsoft Opens the Copilot Harness: 20+ Models, a Rust Runtime, and a Meter That Starts December 1 AI Agents Fragment Exits Stealth With $56 Million So Agents Can Close the Invoice Exceptions Manufacturers Never Finish Chatbots New York Bans 12 Chatbot Behaviors for Minors, With $25,000 Fines and the Power to Delete Trained Models Coding Assistants Cognition Is Fielding Offers at $70 Billion to $80 Billion, Weeks After Raising at $48 Billion News The DOJ Is Asking Whether Nvidia's $12.93 Billion Hugging Face Deal Hands It AI's Front Door Business FuriosaAI Raises $620 Million at $2.9 Billion, Four Times Its 2025 Valuation, With $280 Million From Korea's State Fund News XREAL's Aura Puts Android XR and Gemini Into 95 Grams of Glasses for $1,099 Business AI Hyperscalers Quintupled Flash Prices, and Apple Is Paying $25 Billion for NAND It Once Bought for $6 Billion Coding Assistants Microsoft Opens the Copilot Harness: 20+ Models, a Rust Runtime, and a Meter That Starts December 1 AI Agents Fragment Exits Stealth With $56 Million So Agents Can Close the Invoice Exceptions Manufacturers Never Finish Chatbots New York Bans 12 Chatbot Behaviors for Minors, With $25,000 Fines and the Power to Delete Trained Models Coding Assistants Cognition Is Fielding Offers at $70 Billion to $80 Billion, Weeks After Raising at $48 Billion News The DOJ Is Asking Whether Nvidia's $12.93 Billion Hugging Face Deal Hands It AI's Front Door Business FuriosaAI Raises $620 Million at $2.9 Billion, Four Times Its 2025 Valuation, With $280 Million From Korea's State Fund

Microsoft Opens the Copilot Harness: 20+ Models, a Rust Runtime, and a Meter That Starts December 1

Coding Assistants · By: AIReport Admin · 1 hour ago · 👁 11
Microsoft Opens the Copilot Harness: 20+ Models, a Rust Runtime, and a Meter That Starts December 1

Microsoft put the Copilot harness, the orchestration layer behind Copilot Code, Copilot Studio, the GitHub Copilot app, the CLI and VS Code, into public preview in Foundry Agent Service on October 28. It routes across more than 20 frontier models plus bring-your-own-key, runs on a runtime recently ported to Rust, and is free until December 1, when sandbox compute starts at $0.0994 per vCPU-hour.

Microsoft made the Microsoft Copilot harness available through Microsoft Foundry on October 28, in public preview inside Foundry Agent Service. The harness is the layer the company has standardized across its own agent products: it powers Copilot Code, Copilot Studio, the GitHub Copilot app, the Copilot CLI and VS Code. Until now, outside developers could download pieces of it; from this week they can rent the whole thing, with Microsoft running the agent loop and the sandbox.

The definition Microsoft uses is deliberately narrow. "If the model supplies the intelligence, the harness is everything around the model," wrote Tina Schuchman, corporate vice president of Foundry, and Vladimir Fedorov, chief technology officer of GitHub, in the announcement. In practice that means instructions, working context, an execution environment, policy boundaries and the loop that coordinates model calls and tool calls. The harness supports the Model Context Protocol, agent skills and agent plugins, so a team can change models without rewriting the scaffolding.

Four pieces ship in the preview. Reusable skills package instructions, scripts and supporting resources into procedures an agent can apply across tasks; Microsoft's example is a SKILL.md file that forces a research agent to confirm audience and scope, cite sources and dates, and separate facts from interpretation. Connected tools bring in Foundry tools and MCP servers with explicitly configured authentication and approvals. Execution and delegation let agents touch files and run code and hand bounded subtasks to built-in subagents. And agent definitions can live in source control and deploy through the Azure Developer CLI alongside the rest of an application.

The multi-model claim is the commercial core: more than 20 frontier models plus bring-your-own-key, with Microsoft routing each task to the model or combination of models it judges best. The company also said the runtime was recently ported to Rust, reducing session overhead and time to first token. It published no latency figures and no comparative benchmarks for either claim, so both should be read as directional.

Pricing is where the announcement gets concrete. Harness runtime usage is free until December 1, 2026. From that date, sandbox compute is billed at $0.0994 per vCPU-hour and memory at $0.0118 per GiB-hour, with model inference and any separately billed tools charged on top. Microsoft said prices may vary by region and agreement. For a team running long agent sessions, memory-hours are the number to watch: a 4 GiB sandbox held open for a workday costs more than the CPU it burns.

The preview has real constraints. The harness must be selected when an agent is created, and Microsoft says switching an existing agent to or from it is not supported yet. Foundry manages the sandbox's CPU, memory and idle settings rather than exposing them. Foundry Agent Service now offers three levels of abstraction: prompt agents without the harness, prompt agents with it, and hosted agents where the developer brings the framework and Microsoft only provides hosting. Long-running resilience for hosted agents, which restarts interrupted work from a checkpoint, reached general availability in the same week.

Two customers are quoted. Chris Evans, chief technology officer of Confluence Technologies, said the company had a solution running in less than a week and could spend the time on design instead of infrastructure. Orlando Ribas Fernandes, chief executive of Fashable, described generating tech packs from templates while keeping commercially binding decisions under human control, which is the honest boundary of the product: the harness will run the loop, not take the accountability.

The strategic read is that Microsoft is now selling the part of a coding agent that is not the model. That is a bet that model quality will keep converging while orchestration, governance and sandbox plumbing stay hard, and it moves the lock-in question from whose weights you call to whose runtime holds your context, skills and audit trail. Teams evaluating the preview should measure the exit cost before December, not the free month.

Comments (0)

Log in to join the discussion

Log In

No comments yet