Xum

The trait that decides it is the overview. One screen shows how far each agent has drifted from the branch it started on, which is the number you actually need when three of them have been working for twenty minutes and you have to decide which to keep. Everything else in this category makes you check them one at a time.

It is a terminal tool with terminal ergonomics, including vim-style input, so it rewards habit. Pick it if you live in a shell. Pick Conductor when you would rather click.

The risk is a choice presented as a preference. Running in the project directory means writes land among your uncommitted changes with no separation at all, and nothing in the row records a container boundary to fall back on. The remote option is worse in a different way: an agent executing over a connection to a shared server inherits whatever that account can reach.

What it does right is offering the alternative in the same tool. A local worktree per agent is a first-class mode, not a workaround, and it costs one flag.

1. Separating planning from execution as distinct modes is a sound structure, because it puts a readable artefact between intention and action. 2. Context handling is where the design gets interesting: a manual compaction command coexists with automatic compaction, which means the history a model sees can change without the user asking.

2. That trade, cost against continuity, is never quantified. No token accounting, no benchmark and no description of the trigger, so the behaviour is discovered by watching rather than by reading.

A forced rename is a small event with a useful signal inside it: this shipped fast enough that nobody cleared the name first, which tells you where it sits in the sponsor's priority list. The parent sells remote development environments, and an agent multiplexer that works best over a connection to a server is a demonstration of that product.

Moat: the parent's existing customers, not this tool. Likely path: absorption into the commercial platform as a feature, with the standalone binary maintained as long as it recruits. Position: adopt it, and read the copyleft terms before you embed it in anything you ship.

The remote mode is what makes this interesting to me rather than to individuals: agents run on servers my team provisions, so the compute is centrally managed and the laptops stay quiet. That turns a personal tool into something resembling infrastructure.

Two things hold it back. The licence is strongly copyleft, which is fine for internal use and a conversation if it ever touches a shipped artefact, and nothing here runs unattended, so it produces no scheduled work and no records. Approved with conditions: internal use only, remote workspaces, and a licence note on file.

A single binary is the right shape for something I want on five machines, and copyleft means any fork stays open, which I prefer to permissive for a tool rather than a library. The model configuration treats a local server as an ordinary provider rather than as a compatibility mode, so switching between my own hardware and a routing service is one line.

What is missing is protocol support: no MCP client, so my servers stay outside and the agent's tools are whatever ships. That is the one thing I would patch first.