Version 1.129 Rearchitects How Agents Run Inside the Editor
Microsoft shipped Visual Studio Code 1.129 on July 15, 2026, followed by a 1.129.1 servicing update on July 17 that patched regressions and stabilized the modern UI preview. The headline change is architectural. VS Code now runs AI agent sessions in a dedicated background process called the agent host, and multiple editor windows can connect to the same session. This ends the old model where an agent conversation stayed tied to the single window that launched it. For teams that keep several projects open at once, the shift changes how developers move between a long-running agent task and the rest of their daily work.
The agent host runs agent harnesses including GitHub Copilot, Claude, and Codex, and it sits on a new Agent Host Protocol powered by the Copilot SDK. That combination is the real story for engineering leaders. VS Code is turning agent execution into a shared service inside the editor, with a defined protocol that different harnesses implement. We read this as Microsoft building the plumbing for a future where several vendors' agents coexist in one workspace under a common contract. The immediate benefit is practical, since a developer can start a task in one window, open another project, and keep following the same conversation without losing context.
Why a Separate Agent Process Matters
Running agents in their own process solves problems that grow as agent tasks get longer and more autonomous. A separate host survives window reloads, so a multi-step refactor keeps running while a developer closes and reopens editor windows around it. It also isolates the resource cost of an agent that spans hundreds of files from the responsiveness of the editor itself. For platform teams that standardize developer tooling, this isolation reduces the support burden that comes from agents freezing an entire window. The design mirrors how mature systems separate long-lived work from the user interface that observes it, and it brings that discipline to AI coding.
The multi-window capability changes daily habits in a concrete way. A developer can open the same agent conversation from several windows at once and watch progress from whichever project needs attention. Agents on the host also gain a suite of session-management tools, so an agent can enumerate existing sessions, read their history, and create new sessions for sub-tasks. Sending a message from one chat into another requires user confirmation, which keeps a human in the loop for cross-session actions. We see this as VS Code acknowledging that a single developer now supervises multiple concurrent agents, and the editor needs first-class tools to make that supervision manageable.
The Agent Host Protocol Is a Standardization Play
The most strategic piece of 1.129 is the Agent Host Protocol. By defining a protocol that agent harnesses implement, Microsoft creates a common interface for Copilot, Claude, and Codex to run inside the same host. This echoes the pattern that made the Language Server Protocol so influential, where one contract let many languages plug into many editors. A shared agent contract could do the same for coding agents, giving teams a consistent way to run whichever agent fits a task. For organizations that already mix several AI vendors across their engineering groups, a common host lowers the cost of supporting that diversity across teams.
Standardization also shapes procurement and risk. When agents share a host and a protocol, an engineering organization can swap or add a vendor without retraining developers on a new interface. That optionality strengthens a buyer's position and reduces lock-in to any single agent provider. We would still watch how open the protocol becomes and how much of it stays tied to the Copilot SDK, since a contract controlled by one vendor carries its own dependency. For now, the direction is encouraging for teams that want an editor where multiple agents operate under uniform rules, with session handling and tool access that behave the same regardless of which model sits behind the harness.
Bring Your Own Key, Enterprise Auth, and Portable Prompts
Version 1.129 adds several features that speak directly to enterprise adoption. The Copilot agent harness now supports bring-your-own-key models, so teams can point the agent at model endpoints they control and govern. GitHub Enterprise authentication arrives for Copilot subscriptions, which matters for organizations that manage identity through their own GitHub Enterprise setup. Terminal command execution comes to the chat surface as well, where prefixing a message with an exclamation mark runs a command directly from an agent host session. Each of these changes reduces friction for companies that need agents to fit inside existing identity, model, and command-line boundaries that their platform teams already maintain.
Prompt portability rounds out the release through a migration from prompt files to skills, designed for cross-harness compatibility. This detail deserves attention from platform teams that curate shared prompts across an organization. Skills that travel between harnesses let a team maintain one library and reuse it whether developers run Copilot, Claude, or Codex. We view portable prompt assets as an underrated part of agent standardization, since the value a company builds in its prompt and skill libraries should outlast any single model choice. Combined with bring-your-own-key support, this points toward an editor where the organization owns its agent configuration and carries it across vendors with minimal rework.
What Engineering Leaders Should Take From This Release
For engineering leaders, 1.129 marks a maturity step in how editors host AI agents. The move to a dedicated process, a defined protocol, and session-management tools shows that agent workflows have grown past the experimental phase and now demand real infrastructure. Teams evaluating their AI coding strategy should test the agent host with their own mix of Copilot, Claude, and Codex, and measure how multi-window sessions affect developer flow. The features that touch identity and model control, namely GitHub Enterprise authentication and bring-your-own-key support, deserve early review from security and platform owners who set the guardrails for agent usage across the company.
We also advise care around the experimental surfaces. The redesigned Agents window editor panel and the modern UI preview remain in flux, and the 1.129.1 patch that followed within two days shows how quickly these areas are still moving. Rolling the stable agent host to broad developer populations makes sense now, and pinning the experimental panels to volunteers keeps disruption contained. The larger takeaway is that VS Code is positioning itself as the standard host for multi-agent development. Organizations that shape their tooling and prompt libraries around that model today will find themselves well placed as the Agent Host Protocol and its ecosystem continue to expand.
