Anda

The deciding trait is model tiers with names. Requests route through labelled levels, so sending a trivial classification to a small model and a hard plan to a large one is a label rather than a refactor, and the cost difference over a month is not small.

What you are accepting is a young project in a language where the agent ecosystem is thin, so there are few examples and fewer neighbours. Pick it if Rust is non-negotiable. Pick Koog if you just want a typed language with more people in the room.

The ratio is the concern. A framework this broad implies a maintenance obligation that a project this small cannot obviously meet, and the README does not publish a single install command for the server component, which suggests the getting-started path has had less attention than the architecture document.

What it does right is defaults. Shell access and filesystem reach are optional extensions rather than always-present tools, so an agent starts with a narrow capability set and widens only when somebody chooses to.

1. Every framework that attaches dozens of tools pays for their descriptions on every turn, and almost all of them ignore the cost. Grouping capabilities and deferring schema retrieval until selection converts a fixed per-turn tax into an occasional one.

2. The completion runner accounts usage across nested agent calls, so that saving is measurable rather than asserted. 3. No measurement is published, which is a pity, because this is one of the few designs here whose benefit could be shown with a single token count.

There is nothing commercial to evaluate. No hosted tier, no support offer, no pricing page, and a star count that indicates the work has not yet found an audience beyond people who like Rust. That is a perfectly respectable place for a library to be and a bad place to be depended upon.

Moat: none. Likely path: the good ideas get read and reimplemented in a framework with more users, while this stays a well-designed reference. Position: read the architecture document, borrow the capability-expansion pattern, and do not put a product on top of it this year.

The budget line is zero and the personnel line is the whole cost. Of my sixty developers, the number who would be productive in this codebase is small, and anything built on it becomes owned by that small group permanently, which is a concentration risk I have to manage rather than a tool I can roll out.

It also runs nowhere unattended, so there is no scheduled work, no records and nothing to report on. Support is a repository with a handful of watchers. Not yet: this belongs in a spike, not in an estate.

Signed calls as a first-class context capability is an unusual and welcome detail, because it means outbound requests can carry provenance without me writing the middleware. Each agent gets its own state, cache and object storage, so isolation is structural rather than a convention I have to respect.

The gap is where the model runs. Provider adapters sit behind a shared contract and the row does not record a local option, so my hardware stays idle. Permissive licence, Rust source, so that is a patch rather than a complaint.