Microsoft said on 28 October 2026 that developers can now build their own agents on the Microsoft Copilot harness, the layer that runs Copilot’s agents, through Microsoft Foundry. The harness is in public preview in Foundry Agent Service. Using it is free until 1 December 2026, when consumption-based billing begins.
A harness is everything around the AI model: the instructions, context, tools and controls an agent needs to finish a task. The harness was built by GitHub engineers, and Microsoft has adopted it company-wide, according to a post by Tina Schuchman, Microsoft’s corporate vice president for Foundry, and Vladimir Fedorov, GitHub’s chief technology officer. It works with more than 20 frontier models, plus models customers bring with their own keys. Microsoft reports quarterly results after the US market close the same day.
Short explainer · AI agents brief
A vertical news Short narrated from the published headlines and opening lines of this piece and three related FourWeekMBA pieces on AI agents, models and pricing.
Business Pill · Model Routing
An evergreen Business Engineer explainer on model routing. It explains the general idea only and does not discuss Microsoft’s Copilot harness.
The key insight: As we read it, Microsoft is now offering the agent layer that sits between customers and the models, with a published price for the sandbox, rather than only the models or the finished Copilot apps.
In the Artificial Analysis Coding Agent Index, Copilot Fusion in GitHub Copilot cost $5.59 per task against $11.42 for Claude Code with Opus 5.5 (max), two points lower on the index, Artificial Analysis said. Microsoft links this test as evidence for the harness it now offers in Foundry, where billing starts on 1 December 2026. Sources: Artificial Analysis and Microsoft, 28 Oct 2026.
“If you want to compete, you need to be multi-model,” the Microsoft post opens. It says building a harness that works across many models is “incredibly hard” and that customers can now build on the same harness Microsoft and GitHub ship in their own products without maintaining it.
Where it already runs. The Microsoft Copilot harness powers Copilot Code, Copilot Studio, the GitHub Copilot app, the command-line tool, VS Code and more, the post says.
Models. The harness works with more than 20 frontier models plus bring-your-own-key models, and Microsoft says it can route work to the right model or models for each task.
How to use it. Developers can download a software development kit or run the harness in the cloud, managed by Foundry. Microsoft runs the agent loop and the sandbox; the customer brings instructions, skills, tools and the choice of model, according to the post.
What is in the preview. Reusable skills, connected tools including Model Context Protocol tools, execution with delegation to built-in subagents, and agent definitions kept in source control, Foundry’s developer blog says.
Also announced. General availability of long-running resilience for hosted agents, which helps agents recover from interruptions during long tasks, the developer blog says.
The post says the team recently moved the harness runtime to the Rust programming language, which it says cut session overhead and time to first token.
Cost per task in the Artificial Analysis Coding Agent Index: Claude Code with Opus 5.5 (max) $11.42, Devin Fusion (Opus 5.5 max plus GPT-6 Luna high) $8.00 with a score of 62, and Copilot Fusion in GitHub Copilot $5.59, two points below Opus 5.5 in Claude Code. Microsoft’s 28 Oct post links this test as evidence for the Copilot harness; it measures GitHub Copilot with Fusion, not agents built in Foundry. Source: Artificial Analysis, 28 Oct 2026.
What Was Posted on X
GitHub has released Copilot Fusion in GitHub Copilot, which orchestrates between Anthropic and OpenAI models. It scores 2 points below Opus 5.5 in Claude Code at half the Cost per Task in the Coding Agent Index
GitHub describes Copilot Fusion as a runtime orchestration system, which selects and coordinates models to optimize performance, cost, and latency for each task. Copilot Fusion chooses between a number of models from Anthropic and OpenAI, including Claude Opus 5.5, Claude Sonnet 5.5, GPT-6 Astra, and GPT-6 Sol. We benchmarked its max setting in GitHub Copilot CLI in the Artificial Analysis Coding Agent Index v2.0.
Token costs are charged based on the underlying models used to complete a request.
Key takeaways:
โค Cost-efficient for near-frontier performance: Copilot Fusion in GitHub Copilot costs $5.59 per task, versus $11.42 for Claude Code with Opus 5.5 (max). Devin Fusion with Opus 5.5 (max) + GPT-6 Luna (high) scores 62 at $8.00 per task.
โค Slight lead in Terminal-Bench 4.0: The Coding Agent Index v2.0 weights its three constituent benchmarks equally. Copilot Fusion in GitHub Copilot leads Terminal-Bench 4.0 with a score of 71. It ranks third on FrontierCode in the chart at 51, behind Claude Code with Opus 5.5 (max) at 56 and Devin Fusion with Opus 5.5 (max) + GPT-6 Luna (high) at 54. Its SWE-Atlas-QnA-AA score lags at 57, between GPT-6 Astra in Codex and Gemini 4 Argon in Antigravity CLI.
โค Half the token use of Opus 5.5 in Claude Code: GitHub Copilot Fusion used 7.2M tokens per task in the Coding Agent Index, approximately half of the tokens used by Claude Opus 5.5 in Claude Code, and 30% less than Devin Fusion with Opus 5.5 (max) + GPT-6 Luna (high).
— Artificial Analysis (@ArtificialAnlys) Oct 28 16:14
Artificial Analysis posted that Copilot Fusion in GitHub Copilot, which orchestrates between Anthropic and OpenAI models, scores 2 points below Opus 5.5 in Claude Code at half the cost per task in its Coding Agent Index.
We’re redefining the SDLC for the agentic era.
Lots of GitHub updates this morning at Universe:
โข GitHub Copilot Fusion delivers frontier-level quality, 40% faster and at up to 77% lower cost
โข Our unified Copilot harness is now in Foundry, a managed service for building long-running agents that reason, use tools, and get work done
โข Copilot Projects gives you an orchestration loop to build with multiple agents across repos
And weโre expanding capacity and strengthening GitHubโs architecture to make the home of agentic development more resilient.
Satya Nadella posted that Microsoft is redefining the software development lifecycle for the agentic era, listing GitHub updates from Universe including that GitHub Copilot Fusion delivers frontier-level quality, 40% faster and at up to 77% lower cost, and a unified Copilot harness.
What It Will Cost
Billing for the harness in Foundry Agent Service begins on 1 December 2026, Foundry’s developer blog says. Until then, harness runtime use is free; model inference and separately billed tools are charged as usual.
Sandbox compute. USD $0.0994 per vCPU-hour from 1 December, according to the blog.
Sandbox memory. USD $0.0118 per GiB-hour from 1 December, the blog says, adding that prices may vary by region and agreement.
An example, our arithmetic. A sandbox with 4 vCPUs and 16 GiB of memory running for one hour would cost about $0.59 at those rates, before any model charges.
Two early customers are quoted. Chris Evans, chief technology officer of Confluence Technologies, said: “we had a solution running in less than a week.” Orlando Ribas Fernandes, chief executive of Fashable, said the harness gives his company “a powerful foundation for turning complex sourcing processes into governed, measurable agent workflows.”
The Benchmark Behind the Pitch
Microsoft’s post links to an Artificial Analysis test of Copilot Fusion, GitHub’s model-routing setting, as evidence that the harness performs across model families. In that post, Artificial Analysis said Copilot Fusion in GitHub Copilot scores two points below Opus 5.5 in Claude Code at half the cost per task in its Coding Agent Index.
Cost per task. $5.59 for Copilot Fusion in GitHub Copilot versus $11.42 for Claude Code with Opus 5.5 (max), Artificial Analysis said.
A second router. Devin Fusion with Opus 5.5 (max) plus GPT-6 Luna (high) scored 62 at $8.00 per task, it said.
Tokens. Copilot Fusion used 7.2 million tokens per task, about half of Claude Opus 5.5 in Claude Code, Artificial Analysis said.
Connecting the Coverage
Earlier on 28 October, GitHub made Copilot Fusion generally available and said its Max setting landed within two points of Claude Code with Opus 5.5 Max at about half the average API cost per task, as we reported. That was the routing setting inside GitHub’s products; today’s post offers the harness under it to any Foundry customer. Microsoft AI also shipped its own MAI image and cyber models this week, as we reported on 28 October, and Mustafa Suleyman said its cyber model doubled to 50% of scans, as we reported on 27 October.
Outside Microsoft, SAP described picking the cheapest capable model from more than 70, as we reported, and Mistral began serving DeepSeek beside its own models, as we reported; our same-day look at falling AI prices is here. What’s new: Microsoft is selling the multi-model layer itself, with a price list.
The Structural Read
As we read it, the move turns an internal component into a product: the post says the same harness already runs Copilot Code, Copilot Studio, the GitHub Copilot app, the CLI and VS Code.
The pricing separates the two costs: harness runtime is free until 1 December 2026, after which sandbox compute and memory are billed per hour, while model inference is charged on its own, according to Foundry’s developer blog.
The evidence Microsoft cites is a coding benchmark of GitHub Copilot with Fusion, where Artificial Analysis put the cost per task at $5.59 against $11.42 for Claude Code with Opus 5.5 (max).
Tina Schuchman and Vladimir Fedorov, Microsoft
“If you want to compete, you need to be multi-model”
Three Implications
FOR DEVELOPERS The harness is in public preview in Foundry Agent Service and can be run in the cloud or through a downloadable SDK, with more than 20 frontier models plus bring-your-own-key models, Microsoft says.
FOR FINANCE TEAMS Harness runtime is free until 1 December 2026; from then, sandbox compute is billed at $0.0994 per vCPU-hour and memory at $0.0118 per GiB-hour, the developer blog says, with model charges on top.
FOR MODEL MAKERS The harness routes each task to the right model or models, according to the post, so the customer’s model choice sits inside Microsoft’s layer.
The framework argues that the human’s role relocates as agents take over execution: “As execution delegates, human effort moves to delegation, supervision, review, and integration, and the returns to domain expertise rise.”
As we read it, Microsoft is now selling that harness layer directly: according to its post, Microsoft runs the agent loop and sandbox while the customer brings instructions, skills, tools and model choice. The framework was published in July 2026 and does not discuss Microsoft’s Copilot harness.
What Is Not Established
The posts list which surfaces use the harness but not how many customers or agents run on it. The Artificial Analysis figures describe GitHub Copilot with Fusion in one coding benchmark, not agents built in Foundry. The example sandbox cost is our arithmetic from the published rates and excludes model charges. We did not contact Microsoft or GitHub.
This Week In The AI Supercycle
Execution Fault Lines, Concentrated Demand, and Geopolitical Firewalls
This week’s issue: the AI build is not short of demand for its parts. It is short of cheap money, and its end revenue rests on a narrow base of large, often venture-funded buyers.
Microsoft opened the Copilot harness, built at GitHub and used across its own Copilot products, to developers through Foundry Agent Service in public preview on 28 October 2026. Harness runtime is free until 1 December 2026; after that, sandbox compute and memory are billed at $0.0994 per vCPU-hour and $0.0118 per GiB-hour, on top of model charges.
95,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.
A note on sourcing. The harness details come from Microsoft’s blog post of 28 October 2026 by Tina Schuchman and Vladimir Fedorov; the preview contents, customer quotes and prices from Microsoft Foundry’s developer blog the same day; the benchmark figures from Artificial Analysis’s post on X, which Microsoft’s post links. The sandbox example is our arithmetic. We did not contact Microsoft or GitHub.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
