MoFA

The deciding trait is the bindings. One core is exposed to five other languages, so a Java service and a Swift client can share the same agent implementation instead of each team reimplementing it badly in their own dialect. If your organisation has that shape, this is a genuinely unusual offer.

You are the wrong buyer if everything you own is already Python, because you would be paying the cost of a foreign toolchain for a benefit you do not need. Pick it for a polyglot estate. Pick a Python framework for a Python one.

The kernel is small because the work was moved, not removed. Lifecycle, metadata and task scheduling stay inside; every capability an agent needs to be useful is a plugin. That is a coherent architecture and an expensive starting position, and nothing on the row indicates a library of ready plugins large enough to close the gap.

What it does right is generate schemas from the runtime scripts. Tool definitions are derived from the code that implements them, so the two cannot drift apart.

1. The concurrency model is actors, supplied by an existing library, and distribution is handled by a dataflow runtime borrowed rather than written. Both choices mean the failure modes are already documented outside this project. 2. The agent pattern is ReAct, named explicitly, which lets a reader locate the design in the literature instead of inferring it.

2. Six collaboration modes are enumerated without any guidance on selection, which is the gap: a taxonomy is not a method. No evaluation is published and none is claimed.

291 stars against a scope this large is the number that worries me. Building a kernel, bindings for five languages and a plugin system is a multi-year engineering commitment, and nothing on this row explains who pays for year two. Ambition without a funding story is the most common way good infrastructure stops.

Moat: none commercially. Likely acquirer: none, though a cloud vendor might absorb the dataflow integration. Likely path: it slows when the founding contributors are hired elsewhere. Position: read it for ideas, and do not put it in a critical path.

The isolation story is better than most and it is not what my security team will expect. Compiled extensions run behind a WASM boundary and scripted ones run under resource limits, which is genuine containment without a container, and that distinction will need explaining in a questionnaire.

Beyond that this is not a product I adopt, it is a library my engineers import into a service we operate, so SSO, audit logging and retention are things we build rather than buy. Approved with conditions: as a dependency inside a reviewed service, never as a tool handed to developers.

Scripted plugins are the feature that made me read the rest. A tool can be registered from a script at runtime, so adding a capability does not mean a compile cycle, and the fast iteration loop I usually lose in a compiled language comes back. Compile the ones that matter, script the rest.

A local model backend ships, fronted by an OpenAI-compatible proxy, so the weights can be mine without patching anything. Apache-2.0 keeps the fork clean.