Seamless Takeover Requires Four Transfers, and No Vendor Ships All Four
Picking up an agent's work means four things transfer together: the branch, the environment, the conversation, and the live process state. A June 2026 sweep of ten shipping platforms found that every vendor ships a takeover mechanism and none transfers all four. Most move two or three. A few get close for the solo case, where the same person who started the agent enters the agent's machine. None documents the full team standard this section is built around: a teammate taking over work another teammate initiated, with the state intact and a clean audit record.
This page sets the bar the rest of the Human Takeover pages measure against: the takeover decision, multiplayer takeover, the audit record, and the reverse seam.
Seamless means four things move together
A seamless takeover transfers the branch, the environment, the conversation, and the process state, simultaneously. When a human picks up a teammate's in-progress work, all four come with it: the code, the machine it runs on, the discussion that shaped it, and whatever is live right now. Takeover from an agent should meet the same bar.
- Branch. The commits and working tree, exactly as the agent left them.
- Environment. The toolchain, dependencies, services, and configuration the agent was working in. Not a fresh checkout on your laptop that may or may not reproduce it.
- Conversation. The full prompting history: what was asked, what the agent tried, why it made the choices it made.
- Process state. The live terminal: running servers, half-finished migrations, in-flight test runs, process memory.
Drop any one and the human pays a reconstruction tax. Drop the conversation and they reverse-engineer intent from a diff. Drop the environment and "works on the agent's machine" becomes their first hour. The pattern catalog agentic-patterns.com states the cost plainly, quoting Cursor's Aman Sanger: "if an agent completes 90% of a task in the background but the remaining 10% requires human finesse, a clunky handoff process can negate the benefits of automation."
Most shipping products pick two, sometimes three
The tenth platform, Terragon, shipped branch-and-PR handoff and shut down on February 9, 2026; it is omitted below. Last verified: June 2026.
| Platform | Branch | Environment | Conversation | Process state |
|---|---|---|---|---|
| Claude Code (web) | Yes | No | Yes | No |
| Cursor Cloud Agents | Yes | Partial (remote desktop) | Yes | Remote desktop only |
| OpenAI Codex | Yes | No | Yes | No |
| GitHub Copilot coding agent | Yes | No | Readable, not resumable | No |
| Devin | Yes | Yes (enter Devin's VM) | Stays in Devin's UI | Yes |
| Factory (Droids) | Yes | Yes | Yes | Yes (persistent machine) |
| Imbue Sculptor | Yes | Yes | Yes | Partial |
| OpenHands | Yes | Yes (browser IDE/VNC) | Yes | Yes |
| Google Jules | Yes (PR out) | No | Jules UI only | No |
Two strategies show up in the cells, and each drops something. The products that score well on environment and process state do it by moving the human to the agent's machine; the products that move the work to the human drop the environment. Devin, Factory, OpenHands, and Sculptor give you a seat inside the agent's workspace: Factory's Droid Computers keep "filesystem, configuration, credentials, local services, and even process memory intact." Claude Code's teleport goes the other way: it "fetches and checks out the branch from the cloud session, and loads the full conversation history into your terminal," per the docs, but you land in your local environment, not the cloud VM. Branch-and-PR products like Copilot and Jules hand over code and discard the live context entirely.
Factory and OpenHands document the closest solo answer. The work never moves; the human does. But the strongest vendor cells still stop short of the standard team case: teammate B taking over work teammate A initiated, with attribution preserved when control changes hands. Ramp's in-house Inspect writeup is the public exception on teammate sessions, not a vendor takeover product, and it gets its own treatment in multiplayer takeover.
No shared standard exists yet
The fragmented vocabulary is the tell: "teleport" (Anthropic), "handoff" (Cursor, OpenAI), "take over" (Copilot, Factory), "Pairing Mode" (Imbue), "Use Devin's Machine" (Cognition), "background-to-foreground" (the pattern catalog). No shared term, because nobody has named the requirement.
A few sources get close. Augment Code's handoff guide (May 2026) is the nearest match, stating that it "treats handoffs as a systems design problem for engineering organizations, not a prompt-writing problem for individuals," but it centers on escalation and trust calibration, not on reproducing the agent's working state. AgentBay (Alibaba, December 2025) lets a human "at any moment, seamlessly take over full manual control" of an agent's sandbox, and reports a relative success-rate improvement of more than 48% for agent-plus-human over agent-only. That number is the empirical case for takeover mattering at all. But AgentBay solves the streaming transport, not the branch, conversation, or accountability.
The literature is silent on three more things:
- When to take over. Whether to take over rather than steer.
- The audit record. Where AI authorship ends and human authorship begins.
- The reverse seam. What the agent must learn about what the human changed when it resumes.
Why the bar is four, not two
The four-transfer standard falls out of the architecture. It is the fourth of the requirements team-scale AI engineering demands. When the branch lives in the repo, the environment is a persistent cloud workspace, the conversation is a durable external session log, and the process state is the workspace's live state, all four transfers are the same act: attach to the task. Laptop-centric products fail because two of the four artifacts only exist on a machine the human cannot enter. Solo products fall short when the task needs a teammate or an audit-safe handback. Hold vendors to all four, in the team case.