draive

The deciding trait is what comes back. Generation returns a validated typed object rather than text you then have to interrogate, which sounds like a small preference until you have shipped the third defensive parser and watched it break on a model upgrade. Making the shape a contract moves that whole class of bug to the boundary.

What it asks in return is discipline. You declare types before you get answers, and prototyping is slower for it. Pick it when the thing you are building has to survive a year. Pick something thinner for an experiment.

The risk is what the library declines to own. It composes and orchestrates, and the actual contact with a machine belongs entirely to the tools a developer registers, which means the safety story is whatever those tools happen to be. There is no isolation layer here and no guidance recorded in the row about what a tool is allowed to become.

What it does right is being clear about the boundary. This is a composition layer, sold as one, with no pretence that it also protects you.

1. Treating evaluation as a construct of the framework rather than a testing afterthought is the correct ordering, because it makes a quality check something a workflow contains rather than something a team remembers to run. 2. Guardrails declared alongside the generation they constrain keep the constraint next to the thing constrained, which survives refactoring better than an external policy file.

2. No benchmark is published, and the design makes no capability claim requiring one. The argument here is about construction, and it is made in the right register.

Agency-authored open source has a predictable shape. The code is genuinely good, because it came from delivery work rather than a pitch deck, and it exists to demonstrate competence to buyers of services rather than to earn anything itself. That funds maintenance for exactly as long as the services business finds it useful.

Moat: none in the code; the moat is the consultancy's client list. Likely acquirer: nobody buys this asset separately. Likely path: steady low-volume releases tied to client projects. Position: safe to depend on, with the version pinned.

Observability arriving as part of the design saves a quarter of instrumentation work. Our collectors get fed by the thing itself, so an agent workflow becomes as visible in the dashboard as any other service, and incidents get a timeline rather than a story.

There is no seat price and no console, which resolves procurement and hands governance to my team. Whatever gets built with this is a service we operate, so access control, retention and the on-call rota are ours, and no engineer touches it directly. Approved with conditions: it is a dependency, never a tool we hand out.

Provider substitution done properly is rarer than it should be. Changing where the tokens come from without touching the logic that spends them means routing is a deployment decision instead of a rewrite, and the permissive licence means the whole thing stays forkable if the authors wander off.

The disappointment is the destination list. Weights on my machine are not a supported target, and there is no port for the servers I run, so extension happens in Python inside this project. Good terms, narrow horizons.