Stirrup

The deciding trait is the built-in way for the agent to ask you something. Most frameworks make clarification a thing you engineer afterwards, so the model guesses instead, and you discover the guess three steps later in the output. Having a question arrive at the moment of doubt is worth more than another integration on the feature list.

What you give up is structure. Nothing here decides the shape of your workflow, which is freedom if you have opinions and an empty room if you do not. Pick it if you want to steer.

Defaults decide outcomes. Three execution modes exist and the one that requires no extra package is the one that runs model-written code directly against your filesystem, which means the least careful setup is also the most common one. Anyone following the quickstart has already made that choice without being asked to.

What it does right is offering the other two at all, and documenting which extra brings each of them. The isolation is a package away rather than a fork away, which is the correct distance.

1. Summarising rather than dropping is the better of the two available answers, since discarded turns fail silently while a summary at least records that compression occurred. It remains lossy, and what survives is decided by the same component that decides everything else. 2. Skills are modular units rather than one growing instruction file, which keeps additions from perturbing unrelated behaviour.

2. Nothing measures how much capability survives a compaction, and the authors publish evaluations elsewhere, which makes the omission conspicuous.

The strategic logic is good. An organisation whose reputation rests on evaluating models benefits from owning the scaffold those models get evaluated inside, and giving it away costs them nothing they were selling. The audience is small in absolute terms and precisely the audience they want.

Moat: credibility, which is unusually durable and unusually easy to spend. Likely path: continued maintenance while it serves the parent's research, with no commercial layer ever attached. Position: use it, and understand it exists to support a business that is not this one.

Libraries do not enter my estate as products; they enter as dependencies inside services my team then operates, and the cost is the engineering around it rather than a licence. That is fine, and it is a different budget conversation than procurement expects.

The part needing review is the hosted execution option, which moves source code to an external vendor with its own terms and its own retention. That is a data processing agreement, not a configuration flag. Approved with conditions: local or self-hosted execution only until the third-party option has been through legal.

Routing through LiteLLM instead of hard-coding vendors is the choice that keeps this usable a year from now, because whatever endpoint I stand up next will speak that dialect. My weights, my machine, no negotiation with a provider list somebody else curates.

MCP servers attach as tools, the permissive licence keeps a fork legal, and the extras mechanism means I install the pieces I want rather than a dependency tree I did not ask for. The quickstart does read one specific key from the environment, which I would rather it did not assume.