IBM Bob

The trait that decides it in daily use is continuity between surfaces. The same agent that edits files in the desktop application also answers in a shell, so the mental model you build in the morning still applies when you are scripting in the afternoon. Separate Ask, Agent and Plan modes keep the accidental-edit problem smaller than a single chat box does.

It is also expensive for what it does, and aimed squarely at estates rather than at greenfield. Pick it for a legacy migration with a budget behind it. Pick Cursor for everything else.

The risk is concurrency without a rollback. Several subagents work at once, each in its own mode, and this row records no isolation layer and no version-control capability. That combination means overlapping writes land in one checkout and the only recovery is whatever you committed before you started. A half-finished parallel refactor is the worst state a repository can be in.

What it does right is consent: each spawn is approved individually rather than a single blanket grant at the start of a session, so the fan-out cannot widen silently.

1. The architecture rests on routing each task to an appropriate model. 2. The models are not identified anywhere in the published material, which makes the claim unfalsifiable: a reader cannot check whether the routing is principled, or whether it exists.

2. No benchmark accompanies the modernisation claims, and the documentation describes features rather than an edit-and-verify loop, so how a change is checked before it is offered remains undescribed. An orchestration story with no named components is a diagram, not a method.

Renaming and relaunching an assistant line tells you the first attempt did not convert, and the new packaging tells you what did. Specialised workflow packages sold separately from the subscription mean the margin sits in migration work aimed at estates only this vendor can reach. That is genuine pricing power, because the alternative is a consulting engagement at ten times the figure.

Moat: an installed base nobody else can quote against. Likely path: absorption into a services motion rather than a standalone product line. Position: buy the packages if you have the estate, ignore the assistant if you do not.

Seat maths first. Sixty people at $20 is $1,200 monthly, at $60 it is $3,600, and at $200 it is $12,000, so the tier choice matters more than the vendor choice, and the tier choice depends on consumption I cannot model before a pilot. The trial is thirty days, which is enough for one team and not for a forecast.

The shell form integrates with pipelines, which is the piece that matters to us. Identity and retention terms are not on the pages I can read. Approved with conditions: entry tier, one squad, and contract answers before the second invoice.

The configuration surface is a panel. MCP servers can be added there, which is one real extension point and the only one, and everything upstream of it is sealed: proprietary licence, no bring-your-own key, and the row is explicit that inference does not run on my own machine. So the thing that reads my source is a black box I rent by the month.

Nothing survives the vendor here because nothing is mine. Grudging respect for putting the server list somewhere a human can find it.