Microsoft offers its Copilot harness to Foundry customers in public preview
For the head of model risk at a US bank
Microsoft said on Oct. 28 that it is making the Microsoft Copilot harness, the software it uses to run Copilot products, available through Microsoft Foundry, with the Foundry Agent Service version in public preview.12
Microsoft said building and maintaining a multi-model harness is hard, and that customers can now build on the same harness Microsoft and GitHub ship in their own products without maintaining it.1 The harness was built by GitHub engineers and adopted company-wide, according to the Microsoft blog post by Tina Schuchman and Vladimir Fedorov.1
Microsoft runs the agent loop and sandbox, while the customer brings instructions, skills, tools and model choice, the company said.12
The harness supports 20+ frontier models plus bring your own key, and Microsoft said it can route work to the right model or models based on the task. The sources do not describe how the routing decisions are made or whether customers can see or set them.1
Customers choose a supported Foundry model for their agent, according to the Azure blog post. During the preview, switching the harness after creation is not supported, and Foundry manages the execution environment's CPU, memory and idle settings.2
Billing begins Dec. 1, 2026, with consumption-based pricing, and harness runtime usage is free before that date, according to Microsoft's Azure blog.2
Banks that validate one model before deployment will need to ask which model or models a managed harness uses for each task, and how Microsoft's ongoing harness changes are controlled. Model risk and vendor management teams may need to treat the routing policy and harness updates as part of what they review, or pin a specific model.
Microsoft said in a Foundry blog post that billing for the Copilot harness in Foundry Agent Service, which is in public preview, begins Dec. 1, 2026, with consumption-based pricing.2
Our take. Your inventory approves one model per use case. Microsoft now offers to pick among 20+ models per task and keep changing the harness around them. Validation has to attach to the router, not the model.
Action. Model risk: Require Foundry agent use cases to log which model handled each task and pin approved models before production.
What our records say
Palantir sells data-integration and AI platforms, Foundry and AIP, with its own engineers embedded at the bank to build onboarding, KYC, anti-financial-crime and customer-data workflows on the bank's own data and systems.1 The record in banking is mixed: Customers Bank and Citi are announced, not yet live, while Société Générale uses the suite.2 Citi has not said the Palantir build is live, and Customers Bank is using Foundry and AIP to rebuild deposit onboarding.34
From our records
- Palantir (Foundry and AIP)vendor · banks using it (2) Customers Bank, Citi · alternatives Quantexa, SymphonyAI, NICE Actimize
- Microsoft 365 Copilotvendor · banks using it (21) BofA, Wells Fargo, BNY · alternatives ChatGPT Enterprise, Claude, Gemini Enterprise
- GitHub Copilotvendor · banks using it (9) Goldman, TD, Citizens · alternatives Microsoft Copilot, ChatGPT Enterprise, Claude