Lagent

The trait you notice first is familiarity. If you have written a neural network you already know how this composes, because agents stack the way layers do and the wiring is the interesting part rather than a configuration file. For someone experimenting with arrangements of agents, that shortens the distance between an idea and a running script considerably.

It is a laboratory tool and it feels like one: sparse examples, little hand-holding. Pick it for experiments on open weights. Pick smolagents when you would rather have neighbours than novelty.

The failure mode is privilege inheritance. Tools are ordinary callables invoked in the same interpreter as the caller, and nothing in this row records a containment layer, so a model that chooses badly does so with your credentials, your filesystem and your network. For a library aimed at experiments that is survivable. For anything with a user attached it is not.

What it does right is restraint. The abstraction count is small enough to hold in your head, and a framework you can fully read is a framework you can fully debug.

1. There is one interface between components, a message object, and one composition rule, stacking. That gives a multi-agent system a written topology instead of an emergent one, and a reader can trace which unit produced which output. 2. Borrowing a proven mental model is cheaper than inventing one and easier to teach.

2. No evaluation is published: no task suite, no comparison, no measurement of whether the composition helps. The design is auditable and its effect is not, which is the ordinary state of this category.

This is a distribution asset, not a product. The sponsoring lab wants its own open-weight families to be the path of least resistance for anyone building agents, and giving away the scaffolding is a cheap way to buy that. It means funding is not the risk. Attention is.

Moat: none independent of the parent's model line. Likely path: quiet maintenance while the weights matter, then an archive notice when the lab's focus shifts, which is the normal life cycle for this arrangement. Position: read it for the ideas, and do not let a roadmap you do not control become load-bearing.

Cost across sixty developers is the inference we already buy, and there is no supplier relationship to negotiate, which also means there is no supplier to call. Unattended execution is absent, so nothing here schedules, reports or alerts, and any production use would be a system my team wrote around a library somebody else may stop publishing.

Onboarding a Python engineer is a couple of days. The security questionnaire has no recipient. Not yet, and probably never in this form: this is a research dependency, not a platform.

The local path is real and documented rather than implied. The quick start serves an open-weight model through the packaged vLLM backend, which means the whole loop can run with no account, no key and no egress, and that is a shorter road to offline than most projects manage.

An editable install from a clone is the right default, because I am going to change things. What is missing is protocol: no MCP client, so my existing servers stay outside. I would rather write that adapter than give up the licence.