Agno-Go

The deciding trait is that somebody else already had the design argument. This is a port, so the shape of an agent, a team and a workflow arrives settled rather than invented, and if you have written against the Python original the concepts transfer without a translation table. That is a genuinely pleasant place to start.

The catch is that a port is always second in line. New ideas land upstream first and arrive here when someone has time. Pick it if you need Go and can wait. Pick the Python original if you want to be where the work happens.

The failure mode is drift, and it is structural. Chasing another project's endpoints means the definition of correct lives in a repository with different maintainers and no obligation here. Parity is a moving target measured against work you did not commission, and the gap widens quietly between releases rather than announcing itself.

What it does right is declaring the target. A stated parity goal is at least falsifiable: you can diff the two surfaces and know where you stand, which is more than most reimplementations on this board will let you do.

1. Enumerating the coordination modes and the workflow primitives fixes the composition algebra, so a system's control flow can be read from its construction rather than inferred from transcripts.

2. Exposing the runtime behind an HTTP surface separates the agent definition from its invocation, which permits the two to be tested independently.

3. Guardrails and hooks are structural interception points, not a safety claim, and the row makes no capability assertion requiring a number. 4. No evaluation is published. That is consistent with what is being asserted, which is structure.

This is the hardest position in developer tooling: a volunteer reimplementation shadowing a funded project. The upstream sets the roadmap and captures whatever value the ecosystem creates, and the port absorbs the work without any of the leverage. Sixty-six stars says the audience has not arrived yet either.

Moat: none, and structurally none is available. Likely path: the upstream ships an official Go surface and this becomes a footnote, or the author's interest ends first. Position: use it if the alternative is writing it yourself, and keep the abstraction thin enough to walk away from.

Credit where it is due: a health check and persisted sessions mean this can sit behind a load balancer and be restarted without losing a conversation, which is more operational thinking than most libraries in this category show. That lowers the cost of the service my team ends up owning.

Everything else is missing. No identity integration, no audit record, no retention policy and no administrative surface, so every governance answer is one my engineers build. Nothing runs unattended. Approved with conditions: one team, one service, our own logging, and no direct access for the other fifty-nine.

This checks the boxes I check first. Permissive licence, so a fork is mine outright. MCP is a client capability, so servers I already run become tools without me writing an adapter. Two local serving options appear in the provider list by name, which means my own weights are a supported path rather than a hack somebody demonstrated once.

Compilation to a single binary means I deploy it anywhere without an interpreter. The source is small enough to read in an evening, and that is the property that decides whether I can fix it at midnight.