AI agent platforms took a practical step forward in late October, while customer use of agents is still early. Microsoft released a managed harness that runs the agent loop for you, and DoorDash said third-party AI agents have placed only several hundred orders in total. For most small and midsize businesses, that points to one move: put agents to work on internal, multi-step tasks now, and treat customer-facing agent channels as something to watch.
What did Microsoft announce for AI agents?
Microsoft introduced the Copilot harness in Foundry Agent Service in public preview. It is a managed way to build agents that reason, use tools and complete long-running tasks. Microsoft runs the agent loop and the sandbox. You bring the knowledge, procedures and tools that make the agent yours.
The preview combines four building blocks. Microsoft also made long-running resilience for hosted agents generally available. Foundry detects an interruption and restarts the work with its original input, while your code saves progress so the agent can pick up where it left off.
- Reusable skills: instructions, scripts and resources packaged as procedures an agent can apply across tasks.
- Connected tools: supported Foundry and Model Context Protocol (MCP) tools, with authentication and approvals configured explicitly.
- Execution and delegation: agents can work with files and code and hand bounded tasks to built-in subagents.
- Source-controlled definitions: agent definitions can live alongside your application code.
Why does a managed agent harness matter to a smaller team?
Microsoft’s point is that many teams end up building the same execution machinery. They coordinate model and tool calls, keep working context, manage files and decide how to continue when a task takes more than one step. That work matters, but it rarely sets one company’s agent apart from another’s.
For a business without a large engineering team, that changes where the effort goes. Less of the budget goes to plumbing and more goes to the parts only you can supply: your procedures, your data, the tools the agent may use and the way you judge its results.
The trade-off is dependence on one platform. During the preview, existing agents stay unchanged and you cannot switch an agent to the new harness after you create it. Microsoft also manages the CPU, memory and idle settings of the execution environment.
Are customers actually using AI agents to buy?
Not much yet. On its third-quarter earnings call, DoorDash said it had seen only several hundred orders in total from third-party agents. For comparison, its 2025 report put average deliveries at about 3 million a day. CEO Tony Xu said users still mostly order through the DoorDash app.
DoorDash is still investing. It has its own text-to-order agent, a command-line agent and a chatbot called Ask DoorDash, and it released an MCP integration for corporate ordering from other chatbots. Websites are also still deciding whether to let agents in: TechCrunch notes that Amazon has been blocking personal agents while Walmart welcomes them.
What should you do now, and what can wait?
The useful agent work today is inside the business. Microsoft’s own example is an agent that researches a topic, analyzes supporting data and writes an evidence-backed brief. The same pattern fits operational analysis and internal planning: instructions define the outcome, skills define the method and tools provide the evidence.
Customer-facing agent channels are worth watching, not rushing. Decide your policy on third-party agents visiting your site, and keep your product and ordering data clean so it is easy to expose later.
- Do now: pick one internal workflow with several steps that uses information you already hold.
- Do now: design for interruptions by saving progress at each stage, and check before an agent writes to another system so a restart does not create duplicates.
- Do now: keep people in charge of commitments. One customer quoted by Microsoft, Fashable, says commercially binding decisions remain under human control.
- Wait on: building a dedicated channel for customers’ AI agents until demand shows up in your own traffic.
How much does a managed agent cost to run?
Microsoft says billing for the Copilot harness starts on December 1, 2026, with consumption-based pricing. Before that date the harness runtime is free, but model inference and separately billed tools still cost money.
From December 1, sandbox compute is billed at USD $0.0994 per vCPU-hour and memory at $0.0118 per GiB-hour, and prices may vary by region and agreement. In practice, the model calls and the tools an agent uses will often be the larger part of the bill, so estimate those per task before you scale up.
| Option | Best for | Watch out for |
|---|---|---|
| Copilot harness (preview) | Teams that want Microsoft to run the agent loop and sandbox | Preview limits; you cannot switch an existing agent to it |
| Hosted agents | Teams with their own framework, agent loop or middleware | Your code must save progress and avoid duplicate actions on restart |
| Build it yourself | Full control over every layer | You own coordination, context, recovery and hosting |
What should you do next?
Choose one internal task that takes your team several steps today, such as a weekly operations brief or a research summary. Write down the outcome you want, the method a person follows and the systems it touches, then test it on a managed platform before you commit to a custom build. If you are weighing this for your business, see how we approach AI agent development and LLM integration, or book a meeting with our team.
Frequently asked questions
Is the Microsoft Copilot harness generally available?
No. Microsoft announced it in public preview, with billing starting December 1, 2026. Long-running resilience for hosted agents is the part that became generally available.
Should a small business build an agent for customers to order through?
Probably not yet. DoorDash reported only several hundred orders from third-party agents in total. Internal workflows are a safer place to get value and learn first.
What happens if a long-running agent is interrupted?
With hosted agents in Foundry, the service restarts the work with its original input. Your application decides what progress to save and must check for work already done, such as a report already created.
Sources
- Introducing the Microsoft Copilot harness in Foundry Agent Service, Microsoft Azure, 2026-10-28
- DoorDash said it’s only received ‘several hundred’ orders from AI agents, TechCrunch, 2026-10-30
Researched from the sources above with AI assistance and checked by the Aranv team’s publishing standard.