Microsoft’s November 4, 2026 release of VS Code 1.145 adds an opt-in preview that lets developers group conversations with different coding agents. For a developer or small software team, it offers a way to keep a proposed change beside another agent’s critique before starting implementation.
The preview keeps a coding plan and its critique in separate chats within one grouped session. It does not transfer the first chat’s conversation history to the reviewer. And if both chats use the same folder, they can still change the same files. Grouping is not automatic collaboration or proof that a plan is correct.
Check access and the billing route first
A harness is the software coordinating an AI model and its tools. Selecting a Claude model inside Copilot does not switch to the Claude harness. Microsoft’s harness setup guide distinguishes these access routes:
Claude: use Copilot-routed models or supported provider-billed access. When both are configured, the model picker’s Anthropic or Copilot group determines the billing route.
Codex: enable
chat.agentHost.codexAgent.enabledfor the experimental Agents-window integration. Microsoft lists Copilot Pro+ or ChatGPT sign-in, including a free account, as access routes. This is separate from installing the Codex extension; eligibility does not mean unlimited usage.
On managed machines, Copilot enterprise-managed settings disable the Claude and Codex harnesses, even with separate provider accounts. A managed sandbox setting set to false still counts. User toggles cannot override this restriction, which concerns harnesses rather than provider models inside Copilot. If an agent is missing, run Developer: Policy Diagnostics and contact the responsible administrator.
As of November 4, GitHub lists Copilot Pro+ at US$39 per month; its Student plan excludes third-party agents. That subscription price is not a quote for each review. Copilot-routed agent usage consumes AI credits according to the model and tokens used. Check the selected account’s remaining allowance and spending limit; a second chat is neither automatically free nor necessarily twice the cost.
1. Prepare a plan without starting implementation
Open the Agents window with Chat: Open Agents window from the Command Palette, as described in the Agents-window guide. Choose Copilot and enter /plan followed by the intended outcome, constraints and checks that would establish success.
Read the full plan and save a copy. The planning controls distinguish Implement Plan, which starts execution, from Approve Plan Only, which does not and is offered only in supported sessions. Leave implementation for after the critique.
2. Create a fresh chat with another agent
Use the nested-session controls in the Agents window:
In Settings, enable
chat.agentHost.multipleChatProviders.enabled. Mixed-agent chats are off by default and remain experimental.Hover over the unarchived planning session and select the plus button labelled
New Nested Session, or choose it from the context menu.Select a different available harness above the input before choosing the working folder. Workspace options depend on the harness. Check the reviewer’s model and permissions before sending its first prompt.
Each chat has its own model and permission settings; changing the planner’s settings does not change the reviewer’s.
Do not use /fork as a shortcut for switching agents: a fork keeps the source chat’s harness and configuration. A fresh nested chat gives you a separate conversation to which you supply the plan explicitly.
3. Choose the files and permissions deliberately
For a critique of the plan’s text alone, choose No Workspace where offered and paste the material. Microsoft describes this as a separate temporary folder, not a security boundary. It does not give the reviewer your project by default.
For a review that needs current code, the folder choice matters. The current folder includes unfinished changes. A new Git worktree is a separate checkout starting from the selected branch’s committed files; it does not automatically include uncommitted changes or untracked files. The practical consequence: a reviewer in a new worktree may be assessing an older state than the planner. Supply the missing excerpts or choose the intended committed version.
A worktree separates code changes, not disk or network permissions. The same Agents-window guide says worktree sessions use Allow all without manual approval prompts. Do not choose a worktree expecting it to make the agent read-only.
For folder-based reviews, inspect the selected harness’s effective permissions. Claude offers Request approval to ask before changes; that is not a blanket read-only guarantee. Direct Codex’s Default Permissions permits workspace edits and routine commands. Its Auto-Review preset changes who handles approval requests, not that underlying access. A request to “review only” does not change these controls.
If read-only tools are the priority, Microsoft documents Copilot’s Ask agent as unable to edit files or run state-changing commands. That is an alternative to using a different harness, not the same workflow.
4. Give the reviewer the plan and a bounded task
Paste the plan or attach its saved file, together with the requirements and relevant code. The context picker supports Add Context > Files in supported sessions. Confirm that the attachment appears; if that integration cannot attach the file, paste the relevant text. Include a plan version or date so later revisions are not confused.
The following is a suggested review prompt, not a permission control:
Review this coding plan against the objective, constraints and files below. Do not implement it, edit files or run commands.
Find material omissions, unsupported assumptions, compatibility risks and missing verification steps. For each finding, identify the plan step or supplied evidence, explain the consequence and suggest a correction.
Separate supported findings from questions that need more context. State what you cannot inspect. Do not claim to have tested the proposal.
Objective: [fill in]
Constraints and acceptance checks: [fill in]
Plan and version: [paste]
Relevant files or excerpts: [attach or paste]
Unresolved questions: [fill in]A useful review needs the constraints, not just the proposed solution. For example, if an existing interface must remain compatible, include that requirement and its current definition. Otherwise, a suggestion to redesign it may be irrelevant to the task.
5. Resolve the findings before authorizing changes
Separate the response into corrections to accept, claims to verify and decisions that need human judgment. “This could break compatibility” is not yet a demonstrated defect: ask which interface, requirement or plan step creates the conflict.
Return accepted corrections to the planning chat and revise the saved plan. Resolve material disagreements before choosing an implementation action. Two agents can agree and still be wrong; their answers do not replace checking the eventual code changes and verification results.
For the background to how VS Code organizes agent work, see RohitAI’s August explainer on Agent Host sessions. It covers the architecture; use the current documentation linked above for setup and access rules.