Proliferate

The deciding trait is that every agent runs through its own harness rather than a reimplementation of it. Your prompts, your habits and the quirks you have learned all carry over, and when something behaves oddly you can reproduce it in the terminal instead of wondering whether the wrapper changed something.

What you are accepting is one machine doing five agents' worth of work, with the fan noise to match. Pick it if parallel tasks are your real bottleneck. Pick Warren if you would rather push the work somewhere it can run without you.

Scheduling changes the threat model and the product does not change with it. An agent you watch is constrained by you noticing; an agent that wakes at three in the morning is constrained by whatever the harness allows, which is your shell, your keys and your filesystem. No isolation layer is documented, so the blast radius of a bad nightly run is the machine you work on.

What it does right is the worktree per task. Concurrent branches cannot collide, and a bad run is thrown away by deleting a directory.

1. A shared configuration surface across heterogeneous agents is quietly the most useful property here, because it holds the environment constant. Comparing two agents that were configured separately compares two experiments; comparing two that read the same tool definitions compares the agents. 2. Delegation is scoped, with results collected back to the parent, so a subagent's context is bounded by construction rather than by prompt discipline.

2. No evaluation is published, and the design claims are structural, which is the consistent pairing.

A company shipping a free client with no visible commercial layer is either pre-monetisation or pre-decision, and from outside those look identical. The copyleft terms are the tell: they keep a competitor from shipping a closed derivative, which is what you do when you intend to sell something adjacent later.

Moat: the workflow habit, if teams build schedules they do not want to rewrite. Likely acquirer: a vendor that owns one of the five agents and wants the surface around it. Position: use it, and expect a paid tier to appear where the automation is.

Platform coverage decides this before anything else does. A tool that runs on one operating system splits the organisation into people who have it and people who file tickets about not having it, and I have spent years removing that pattern rather than adding it.

The money is not the client, which is free, but the underlying agent subscriptions it drives, already sixty lines on an invoice and unchanged by adopting this. There is no directory integration and no console. Not yet: revisit when the other platforms ship.

Copyleft on a desktop client is a stronger promise than a permissive licence would be here, because it means any hosted derivative owes the source back. MCP servers I already run attach once and every agent inherits them, which removes the tedium of configuring the same three tools five times.

Browser and computer control sit in the same shared layer, so an agent that needs a page gets one without me wiring a bespoke tool. Credentials stay mine throughout, which is the part I check first and the part most products fail.