Microsoft lets AI agents share one runtime with Copilot SDK 2.0
The release adds per-user credentials and resumable sessions. Its faster shared setup leaves application developers responsible for enforcing access boundaries.
Loading page…
The release adds per-user credentials and resumable sessions. Its faster shared setup leaves application developers responsible for enforcing access boundaries.
Listen to this story
Copilot SDK 2.0 lets developers separate the agent’s model loop from the environment that runs its commands, so ordinary sessions can share an application runtime while higher-risk execution remains sandboxed. Durable session journals also let another service instance resume a conversation after checking caller ownership. The practical shift is more control over when to incur sandbox startup and hosting costs—not a removal of isolation requirements for agents that run untrusted code.
On Microsoft’s warmed one-turn benchmark, the Rust core took 55.3 milliseconds versus 5.25 seconds for the TypeScript CLI; the comparison is specific to that workload.
Developers can start sessions in shared hosting and move them into a sandbox when needed, but that transfer is one-way.
The SDK supports TypeScript, Python, Go, .NET, Java and Rust, and can store session state through an application-provided provider.
Microsoft’s Copilot SDK 2.0 promises lighter, faster AI agents by letting many sessions share one runtime instead of requiring a container each. Announced at Ignite on November 17, 2026, the release pairs that change with a Rust core and striking company benchmarks. But it does not eliminate the need to isolate agents that execute untrusted code.
The developer toolkit, previously called GitHub Copilot SDK, gives applications programmatic access to the software that coordinates Copilot agents. It supports coding, non-coding and creative applications. Version 2.0 separates decisions that previously came bundled together: where the agent’s model loop runs, and where it handles files or executes commands.
Microsoft says creating a client and session, then completing one turn, took 55.3 milliseconds with the Rust core running inside the application. The TypeScript CLI took 5.25 seconds on the same warmed workload. That comparison covers two runtime setups under a specific condition—not a promise that every agent task will finish in 55.3 milliseconds.
The engineering history helps explain the change. In a September technical account, GitHub engineer Stephen Toub described an SDK layered on top of the command-line interface. Applications launched that interface as a separate, headless process, carrying Node.js and V8 with it. Calls and session-file operations then crossed a process boundary rather than staying inside the application.
Version 2.0 runs inside the application by default. It exposes a C ABI—a language-neutral calling interface—but developers can still use a child process or shared service. Microsoft says the agent code stays the same across those hosting choices.
In the shared setup, each session keeps its own state, credentials, tools, files and history. Starting another session does not require another container, process or virtual machine. Microsoft says that cuts infrastructure waits and operating costs. But without a sandbox, the application—not a container boundary—enforces isolation. Microsoft describes this arrangement as useful for some assistants, but limited for coding work.
The SDK’s empty mode starts without a host shell, host file-system access or default tools. Developers explicitly add permitted tools and set an allow-list for each session. Credentials also move from the process to the session: model and tool calls can run as the requesting user, rather than as the server. A token-provider callback supports refreshing short-lived credentials.
Durable sessions are the other major change. A session journal records messages, tool calls and results, plans and checkpoints as an event log. Any runtime that can read that log can resume the conversation with its history intact. The default storage is the runtime’s own disk; an application-supplied sessionFs provider can instead place state and session files in object storage, a database or tenant-controlled paths.
That allows a web client to disconnect while a long task continues, or a webhook to restart a conversation hours later. Another service instance can resume a session by ID after checking that the caller owns it. For multiple clients attached to one live session, the SDK can sit behind an Agent Host Protocol host, which synchronizes their access.
The SDK is available in TypeScript, Python, Go, .NET, Java and Rust. Developers building these architectures run them on their own infrastructure today. The release gives them more control over when to pay the startup and hosting cost of a sandbox, rather than requiring one for every session.
Loading discussion...
Join the conversation
Explain when stronger separation is worth the extra cost.
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.