tech-news · 05 November 2026
Microsoft’s Intelligent Terminal 0.3 Makes Agent Work Visible
Intelligent Terminal 0.3 adds live agent status, cross-window session search and headless tabs. The useful shift is operational visibility, with important limits.
Microsoft released Intelligent Terminal 0.3 on 4 November, turning its experimental Windows Terminal fork into something closer to a control surface for command-line agents. The update does not introduce a new model or promise autonomous software delivery. It tackles a less glamorous problem that appears as soon as several agents are working at once: knowing which session is active, which one needs attention, where each one is working, and whether closing a window will kill a long-running job.
That matters now because agent tooling is moving beyond one chat beside one editor. Developers increasingly run separate sessions for implementation, review, testing and investigation. Ordinary terminal tabs were designed to identify shells, not to expose the state and history of semi-autonomous processes. Intelligent Terminal 0.3 is an early attempt to make that work legible.
What changed this week
The centrepiece is a vertical sidebar that shows shells, split panes and agent sessions together. According to Microsoft’s release announcement , each agent tab can display live status and an attention marker when human input is required. Tabs can also show selected context such as the working directory, Git repository, branch or change state. Pinning, an agents-only filter and global search make the view useful when the session count grows beyond what a horizontal tab strip can explain.
Search crosses more than the current window. The recent-sessions list includes agents running in other Intelligent Terminal windows as well as sessions that have already closed. Selecting an entry can focus a known live pane, restore a retained tab, or start a supported resume flow. That last qualification is important: resumption still depends on what the underlying agent CLI can do. The terminal can catalogue a session, but it cannot manufacture durable session semantics for every tool.
Version 0.3 also discovers installed agent CLIs and creates new-tab profiles for Copilot, Claude, Codex, Gemini, OpenCode and Antigravity. These profiles launch the tools’ own interactive interfaces. They do not merge those products into one agent runtime, and each tool keeps its own authentication, subscription and permission model. The automatic profiles currently cover Windows installations, not equivalent CLIs installed inside WSL.
The most immediately useful addition may be headless mode. A selected tab and its existing panes can remain alive after the tab or terminal window is closed, with retained work available from the sidebar or system tray. The independent WindowsForum account highlights the crucial boundary in the detailed release notes: this is live-process retention, not a service. It does not survive Intelligent Terminal exiting or Windows rebooting. Startup restoration can reconstruct supported layouts and conversations, but it does not preserve the old processes.
Why it matters
The interesting change is not the sidebar itself. It is the recognition that agent work needs an operational layer.
A conventional terminal assumes the person remembers why every tab exists. Agent workflows weaken that assumption. A session may run for tens of minutes, pause for approval, or finish while attention is elsewhere. Titles such as “PowerShell” or “copilot” are poor operational signals. Showing status, repository context and attention state in one place reduces the need to click through every pane simply to discover what is happening.
Headless retention addresses another common failure mode. Closing a window is a user-interface action, but terminal-hosted work often treats it as process termination. Letting a selected tab outlive the visible window is a practical convenience for builds and bounded agent tasks. Microsoft has also kept the limitation understandable: it is not durable execution. Work that must survive crashes, updates or reboots still belongs in a proper background service, scheduler, CI runner, container supervisor, tmux -style environment, or another system designed for recovery.
There is a wider design lesson here for agent platforms. Visibility and control should live outside the agent’s own narrative. A model can say it is testing, waiting or finished, but the host should independently know whether the process exists, what workspace it occupies and whether it is requesting input. Intelligent Terminal is not a full orchestration system, yet its status signals, session index and explicit retention controls point in that direction.
The trade-offs are real. The project remains experimental, Windows-specific and coupled to the capabilities of each supported CLI. A sidebar consumes useful screen width. Recent-session discovery is not the same as guaranteed recovery. Running several agents also increases the chance that concurrent sessions touch the same repository or compete for credentials and local resources. Better visibility helps operators spot that risk; it does not provide workspace isolation, change coordination or policy enforcement.
What to watch next
The next evidence should come from reliability rather than a longer provider list. Watch whether live status remains accurate when agents crash, detach or wait on hidden prompts. Session restoration needs testing across terminal updates and forced Windows restarts, especially where the distinction between reconstructed conversations and dead processes may confuse users.
It will also be worth watching whether Microsoft adds stronger workspace boundaries. Repository locks, clear collision warnings, per-agent resource information and explicit permission state would move the product from session navigation toward genuine local-agent operations. WSL parity matters too, because many Windows development workflows keep their actual toolchains and agent CLIs inside Linux environments.
For now, Intelligent Terminal 0.3 is best read as a useful interface experiment. It does not make agents more capable. It makes several running agents easier to observe, find and deliberately keep alive. As agent tooling matures, that operational clarity may prove as important as another jump in model intelligence.