AgenticGoKit

The deciding trait is language, and it decides more than people expect. Keeping agent code in the same binary as the service it belongs to removes a deployment, a container image and a whole class of packaging argument. If your team already ships Go, that saving is real on day one and it keeps paying.

What you trade is everyone else's homework. The Python side of this category has years of tutorials and copyable patterns; here you read the source and work it out. Pick it if the language matters more than the head start. Pick Python if it does not.

The risk is interface churn, and the project states it plainly. A surface named v1beta is a surface reserving the right to move, and orchestration frameworks are the worst place to absorb that, because the agent graph is not one call site but the shape of your whole system. A rename in the streaming interface is a refactor across every handler.

Two things it does right. The warning is on the tin rather than in a changelog nobody reads, and the error handling was written as part of the release rather than added after the first support thread.

1. Sequential, parallel, DAG and loop are enumerated composition operators, so a topology is a declaration rather than a property inferred from prompts at runtime. That is the correct level to expose. 2. Tracing is native rather than an add-on, which means a run leaves a spans record and a latency question has an answer instead of an opinion.

2. No evaluation is published, and none is claimed, which is consistent. 4. The multimodal path accepts images, audio and video, but the row demonstrates ingestion rather than any measured capability over it.

There is nothing to underwrite here. No entity, no revenue line, no hosted product to convert the users into anything, and a domain paid for by somebody who wanted the docs to look serious. That is not a criticism of the code, it is a description of the risk you are taking.

Moat: none. Distribution is a package import in a language whose users are famously willing to write it themselves, which is the weakest possible position. Likely path: absorbed as a pattern by a larger project while this one slows. Position: fine as a dependency, wrong as a foundation.

Procurement never sees this, which is convenient until an incident does. No console means no central model configuration, no key rotation policy I can enforce and no place to read what an agent did last Tuesday. The permissive licence makes legal review a formality.

Nothing in the row runs unattended, so it is not a pipeline component and never appears in a throughput number I can report. Onboarding a Go engineer is short; onboarding anyone else is not. Not yet as anything my organisation adopts, though I will not block a team using it inside a service they already run.

The model story is the part I like. Ollama and vLLM sit alongside the hosted providers rather than in a contributed corner, so pointing the whole thing at my own box is a config change and not a fork. Any key I already hold works. The licence keeps a fork alive if the author disappears.

The protocol gap annoys me. No MCP client means the servers I run are not tools here until I write the bridge myself, and in a framework that is my code to maintain forever. Static binaries buy back some goodwill.