Feedback wanted: keeping an embeddable agent core free of app dependencies in Go

Technical Discussion
1

Hi all,

I maintain suga, a coding agent written in Go (MIT, still v0). Most of the work has gone into long sessions, but the part I would like Go feedback on is how the core is split out so other Go programs can embed it.

pkg/agentcore holds the model/tool loop, providers, the tool registry and approval. It must not import anything under internal/. Instead of relying on review, a test walks the package imports and checks them against an allowlist, so a new dependency or an unregistered package fails go test. The smallest working setup is:

runtime := agentcore.New(agentcore.Config{
    Provider: mock.NewProvider(mock.Final("hello")),
})
result, err := runtime.Run(ctx, agentcore.RunInput{
    AgentID:     "example-agent",
    Instruction: "say hello",
})

Wrapping the registry in tools.NewApproval with no approver denies every call, so the default is closed rather than open.

Two things I am unsure about:

  • Is an import allowlist test a reasonable way to keep a layer boundary, or would you split the core into its own module?
  • The in-process SDK (pkg/sugasdk) is still marked experimental. For embedding an agent loop, would you rather get a callback-style API or a channel of events?

Disclosure: I’m the author, and I use AI tools heavily on this project, including for drafting this post.