Visual Studio Code compresses agent state

Microsoft’s May 20 release of Visual Studio Code 1.121 is partly a product update and partly a statement about how coding agents should be run. The release notes say the editor now uses “more output compression” and “background terminal cleanup” to cut resource use and token load in agent workflows. They also add handling for sensitive terminal prompts so secrets stay in the terminal rather than getting pulled into agent context. (code.visualstudio.com) Those are small-sounding changes, but they target three recurring failure modes in agent systems. First, context bloat. Terminal-heavy agents can dump large volumes of logs, test output and install traces into the working context. VS Code says 1.121 streamlines terminal tool behavior with more output compression, which means the tool is trying to preserve signal without replaying every line back into the model. That fits Microsoft’s broader documentation around “context engineering,” where VS Code frames context as something that should be curated rather than accumulated. (code.visualstudio.com) Second, stale side effects. The same release says terminal behavior now includes background cleanup. In practice, that points to a familiar problem in agent sessions: shells, watchers, servers and child processes outliving the task that spawned them. VS Code does not present this as a headline feature for users so much as an operational safeguard for agent runs, especially as its Agents window expands and remote agent sessions keep running on long-lived “agent host” processes. (code.visualstudio.com) Third, prompt and secret isolation. VS Code’s release notes say sensitive prompts such as passwords, passphrases, PINs and verification codes are now intercepted when a chat terminal command reaches them, because an agent could otherwise capture or replay secrets. That matters because modern coding agents increasingly mix terminal access, chat history, memory and reusable prompt assets in one workflow. Microsoft’s docs show that prompt files, custom instructions and memory all persist in different ways, with some memory loaded automatically at session start and some scoped only to a session. Keeping sensitive prompts out of that broader context reduces the chance that secrets or privileged instructions bleed into visible chat state. (code.visualstudio.com) The release also lands as VS Code pushes deeper into agent-specific tooling. The 1.121 notes highlight the Agents window, remote agents in preview, and model configurability for lighter-weight tasks. Separately, Microsoft’s documentation now breaks customization into prompt files, custom agents, custom instructions and memory, each with different persistence and control boundaries. That architecture suggests the company is treating agent behavior less like a single chat box and more like a system with separate layers for task prompts, durable preferences, temporary plans and tool access. (code.visualstudio.com) For enterprise teams, the practical takeaway is straightforward. The hard problems in agent deployment are often not model quality alone but what gets remembered, what gets compressed, what keeps running after the task ends, and what never should have entered shared context in the first place. VS Code 1.121 does not solve those problems in general, but it shows them being addressed at the tooling layer, inside a mainstream development environment. (code.visualstudio.com)