Pimp My IDEBack to the garage
October 7, 2026 / remote agents / control paths

A remote control does not move the engine.

Agent tools now let you steer work from a phone, a browser, or another editor window. That convenience does not tell you where commands run, which files are present, or who can stop the job.

Name the steering path and the execution host separately.
Execution hostFILES / TOOLS / CREDENTIALS

The phone is a steering surface.

Cursor's October 6 release lets its iOS app show and reply to local agents running on a paired computer. The agent stays on that computer. Cursor says the computer must remain on and online. Its optional keep-awake setting also requires power and an open lid.[1]

That is remote steering, not remote execution. The phone carries the conversation. The computer still owns the files, tools, processes, and failure state. A dead battery, closed lid, lost network route, or stuck local process can end the useful path while the mobile interface remains perfectly understandable.

Do not write "remote" on the run card. Write who steers, where tools execute, and what survives a disconnect.

A remote host changes the machine.

VS Code's remote agent sessions put the Agent Host on another machine and connect over SSH or a dev tunnel. The machine must be powered on and reachable. The browser or desktop window is a client to that host.[2][3]

Its experimental delegation tools make the missing contracts explicit. They can select a named host or match operating system, memory, CPU, and model requirements. They do not clone or copy the coordinating workspace. A session starts without a workspace unless the request names one. The final response is not forwarded by default, so the delegated agent must send a message back while the coordinating window remains connected.[2]

Placement is one decision. Workspace, authority, and return are separate decisions. Automatic host selection cannot repair a missing revision, an overpowered credential, or a result with nowhere to land.

Self-hosted can still split control.

Cursor's Self-Hosted Machines put file edits, terminal commands, computer use, and local MCP servers on a machine the customer manages. Cursor keeps the agent loop, inference, and planning. The worker sends needed file contents, terminal output, diffs, screenshots, local MCP results, and routing metadata to Cursor during a run.[4]

This is a useful custody option, but it is not an offline agent. The full checkout and machine-local credentials can stay on the worker while selected content and results cross the outbound connection. The docs say no inbound port is required. They also tell operators to keep secrets out of tool output and uploaded artifacts.[4]

Write the dispatch card before launch.

Name the host and its wake condition. Name the exact folder, revision, and worktree. List approvals, network access, credentials, merge rights, and the stop owner. Define how status and the final artifact return after a client disconnects.

A mobile control can be the right tool. A remote host can be the right tool. A self-hosted worker can be the right tool. None of them replaces an execution map.

Interactive makeover / remote route mapper

Remote steering breakaway

Traditional purpose replaced: one vague local-or-remote toggle. Better version: one native route selector plus four independent contracts. The control writes a dispatch card without claiming that a connection, approval, or run exists.

Select the operating route
Add contracts to the dispatch card
Phone to local0 of 4 contracts

No contracts are selected. The route is a label, not a dispatch plan.

No contracts are selected. Evidence is still required.

Sources read / October 7, 2026

Source log

Evidence boundary. These sources describe vendor products and their documented control paths. This dispatch did not pair an iPhone, start a VS Code remote host, or register a Cursor worker. Product availability, defaults, and preview status can change. The breakaway maps required contracts. It does not test a connection or grant authority.