Summary: On October 26, 2026, GitHub announced that assisted permissions are generally available in Copilot CLI, the Copilot app, and Visual Studio Code. An LLM now evaluates each permission request — commands, file paths, URLs — and clearly safe ones run without a manual prompt. For teams running agentic coding workflows, this cuts approval fatigue while keeping enterprise policy in charge.
How the evaluation works
Instead of asking you for every single tool call, Copilot routes each request through a classifier. Requests that the model judges clearly safe proceed automatically; everything else still waits for your approval. That’s a middle ground between the two extremes most of us have lived with: approving trivially safe commands dozens of times a session, or blanket-allowing everything and hoping for the best.
The mechanism is the same across the three surfaces, which matters if your team works across them. Per the changelog: run /permissions and select assisted in Copilot CLI; open the mode picker, choose Tool permissions, then Assisted in the Copilot app; or select Assisted permissions from the permissions picker in VS Code.
What administrators get
The enterprise controls are the part worth reading closely if you manage a team. Admins can block Allow All while keeping Assisted available — so developers get automation, but never a blanket yes to everything. You can also disable Assisted separately if your organization would rather keep approvals fully manual.
Admins can set the default permission mode for new sessions, though developers can still pick another mode that policy permits. If the configured default is unavailable or prohibited, the session falls back to manual approval. That fallback is a sensible design: the failure mode is more friction, not less oversight.
Where this fits in your workflow
Approval prompts are one of the quiet costs of agent-assisted coding. In my earlier write-up on Copilot’s diff, terminal, and browser review loop, the tradeoff was visibility versus interruptions — every guardrail you add costs a click. Assisted permissions attack the same problem from the permission side: fewer prompts for genuinely safe operations, with human judgment reserved for the ambiguous cases.
The practical question for builders is what “clearly safe” means in practice for your codebase. GitHub points to its documentation on allowing safe commands without manual approval for the evaluation details, and it’s worth reviewing that before rolling this out — especially if your workflows touch production systems, credentials, or anything where a file-system write is not a routine event.
A grounded next step
If you run Copilot across a team, the reasonable move is to enable Assisted in a low-risk environment first, watch which requests get auto-approved, and then decide whether your enterprise settings should pin the default or keep Allow All blocked. The feature is GA as of the October 26, 2026 changelog entry, so it’s ready for real use — the judgment call about what safe means for your repo is still yours.
Sources
View the original post on X