Visual Studio Code stable releases v1.120 through v1.123, shipped throughout May and early June 2026, move Copilot agent features into a preview — ready Agents window and roll out several agent, model and terminal improvements. The change centralizes an agent — first workspace for cross — project tasks and introduces capabilities — like persistent remote agent sessions and expanded BYOK options — that directly affect developer workflows and enterprise deployments.
The new Agents window provides a focused surface for completing tasks across multiple projects, with faster navigation and change review. Remote agents (preview) let sessions run on remote machines over SSH or Dev Tunnels and keep those sessions running when the client disconnects, enabling longer — running workflows. Git flow is tightened for agents: the environment pulls base-branch updates before an agent begins edits and automatically refreshes Git state after commits and syncs to reduce merge surprises during automated edits.
Deployment and model control options were expanded with teams and air-gapped environments in mind. BYOK support is extended to air-gapped setups, a model picker is now organized by provider, and a Custom Endpoint provider option enables endpoints compatible with chat completions, responses, or messages. Session sync to users’ accounts creates a searchable history across machines, improving auditability and support for distributed workflows.
The updates change how builders interact with agents day to day. Agent workflows can survive client disconnects, recent preferences such as harness and isolation modes persist across sessions, and developers can run side‑by‑side agent sessions for parallel review. Chat sessions now sync automatically to users’ accounts, and a /chronicle command makes it possible to query past work to generate standups, summaries or productivity tips from prior sessions.
Language model controls and observability were strengthened. BYOK token visibility now reports actual token usage for bring‑your‑own‑key models, while new reasoning‑effort controls let teams trade quality, latency and cost directly from the model picker. Administrators can also configure which models handle utility tasks — titles, summaries, rename suggestions, commit messages and intent detection — so different models are assigned to appropriate workloads.
Terminal and tooling changes aim to reduce token consumption and improve safety. Expanded terminal output compression trims verbose build and test output passed to models; sandboxed, network‑dependent commands are automatically retried with broader network permissions while preserving filesystem protections; and an experimental command risk assessment adds AI‑generated risk labels and short safety explanations. The VSCODE_AGENT environment variable enables CLIs to adapt to agent‑initiated commands, and the integrated browser gains device emulation, selectable screenshot capture attachable to chats, saved favorites and local HTML file preview to help reproduce UI issues.
Sources
Replies (0)
No replies in this topic yet.
