HyperlightCodeActProvider — alpha feedback

[eavanvalkenburg]

We just released agent-framework-hyperlight, an alpha package that
brings CodeAct-style tool calling to Agent Framework on top of
Hyperlight. The short
version: the model gets one execute_code tool, writes a small Python
program per turn, and calls your tools from inside a fast, locally
isolated micro-VM via call_tool(...). Full write-up: <link to
blog post>.

Before we harden and promote the package, we would like feedback on
four deliberate limitations in this first cut:

1. Platform coverage. The backend currently ships for Linux and
   Windows. macOS support is on the roadmap. If macOS is blocking you,
   say so here — it helps us prioritize.
2. Language coverage. Only the Python Agent Framework is
   supported today. A .NET counterpart is coming soon.
3. Guest runtime. We ship a Python guest. Hyperlight can host
   other guests (JavaScript in particular), but each new guest means
   a new CodeAct instruction template, a new call_tool surface, and
   new ergonomics to maintain. Would a JavaScript guest (or another)
   actually change what you could build? If so, describe the scenario.
4. Approval model. Approvals currently apply to the whole
   execute_code call, not to individual call_tool(...)
   invocations within it. Per-operation gating requires keeping that
   operation as a direct agent tool. We are considering richer
   options — per-tool prompts from inside the sandbox, post-hoc audit
   hooks, policy-driven approval. Which of those would you actually
   use, and for what?

Please reply with:

• Your scenario (what the agent is doing, how often, how autonomous).
• Which of the four items above, if any, blocks or complicates it.
• Anything we have not listed that is in your way.

Bug reports and concrete feature requests are better filed as issues
at
github.com/microsoft/agent-framework/issues
so they do not get lost in the thread — this discussion is for
directional feedback on the limitations above.