Agent session portability: start a task on one agent, finish with another

Switching coding agents halfway through a task means bringing the next agent up to speed. The repository contains the code, but the conversation holds context about what we asked for and where the work stopped. When a rate limit interrupts a session, reconstructing that context becomes another task to finish before coding can continue.

Session portability in the Speakeasy AI Control Plane prepares that handoff from the recorded session. Choose Cursor, Codex, or Claude Code, and Device Agent opens it in the same repository with the task and recent conversation ready to continue. The handoff is assembled on the machine without uploading the session content.

How coding agent session portability works

Claude Code, Cursor, and Codex each keep a local transcript of a session on disk: the messages exchanged, the tool calls made, and the files touched. Device Agent reads that transcript and extracts five things from it: the original task, the most recent turns of conversation, the list of files the session touched, the last message the assistant sent, and an instruction to continue. It writes those into a single handoff prompt.

The target agent then starts a new conversation and receives that prompt as its first message. This is not the same as importing history. The new session has no memory of the old one beyond what the handoff prompt says. It reconstructs enough context to continue the task, not a replay of the conversation that produced it.

In plain English, Device Agent is not moving a session file from one tool to another. Coding tools store sessions in incompatible formats, so there is no shared format to move. Instead, it reads what is already on disk, summarizes the parts that matter for continuing the work, and hands that summary to the next agent as a normal prompt. Any tool that accepts a prompt can be a target, which is why this works across three unrelated products.

This remains a user-steered agent session. The handoff changes the client that carries the work, not the person whose identity and permissions govern the next tool call.

How to move a coding agent session

This release is available in the macOS Device Agent public beta. Before starting, confirm that Device Agent is enrolled on the machine and Session Portability is enabled for your organization. The target coding tool must also be installed on that machine.

Continue from the menu bar

1. Open the Device Agent menu bar and choose "Continue session".
2. Pick a recent session. Each entry shows the task, tool, time, and repository to help identify it.
3. Choose the coding tool to continue in.

The target opens in the session's project directory with a handoff that carries:

• The task.
• Recent turns.
• Files touched.
• The last assistant message.
• An instruction to continue.

Continue from the CLI

List the recorded sessions, then replace <id> with the session ID to move and run the second command to continue in Cursor.

speakeasy sessions list
speakeasy sessions move <id> --to cursor

How the handoff arrives depends on the target. Claude Code and Codex start a new session with it as the first prompt. Cursor has no import API, so the workspace opens with the chat prefilled from SESSION_HANDOFF.md. If a launch fails, the files are already written and the exact command is copied to the clipboard.

What stays local and what gets logged

A move stays on the machine. Device Agent reads the transcripts the tools already keep on disk, assembles the handoff locally, and uploads no session content. Each move writes a content-free audit event that records who moved which session, into which tool, and from which device. Security teams can see that the move happened without seeing the conversation.

For help rolling out Device Agent across a team using different coding tools, book time with our team.