Back on The Cloud Gambit: MCP, NetClaw, and AI Governance

I’m back with William Collins on The Cloud Gambit, part of the Packet Pushers network, for Why Training AI to Have a Self Makes Governance Harder.

William and I have been talking about MCP on this show since early 2025. This time, I had just returned from MCP Dev Summit Toronto, and we had plenty to discuss: what it takes to connect agents to infrastructure, how I use a constitution in NetClaw, and what happens when the model’s own training introduces expectations the operator did not write.

From the summit to the network

At 2:31, we get into Toronto and the distance between enthusiasm for agents and willingness to let one change a network. I describe asking the room who would ride in a self-driving car, then who would let an agent make a network change. That second question still produces a very different response.

I also revisit NetClaw’s beginnings in pyATS and MCP. The motivation was to make network engineering knowledge and tools easier to use through natural language. If you want the longer version of that journey, my Toronto story, photos, and slides are here.

At 8:34, I explain why I value exposing a defined MCP tool: I can identify the operation, log its use, and apply controls around it. We talk about skills and MCP working together, and the summit’s emphasis on implementation, policy, and governed access.

A constitution starts with operational expectations

William turns to NetClaw’s constitution at 17:54. I explain how the project grew through iterative agent work and how specification-driven development helped me make its expectations explicit.

The practical examples begin around 22:12: establish the health of the environment before a change, and require the appropriate approval before changing it. Those ideas come from my experience operating networks and managing change windows. They belong in the agent’s operating instructions too.

We also talk about the inconvenience of your own rules applying to you during a demo. That is a useful moment to examine the rule, understand its purpose, and amend it deliberately when the requirements change.

At 25:53, we move from the constitution to the specification and clarification step. A request for a multiplayer game still leaves unanswered questions about how many people should be able to play. Human input at that stage can prevent an agent from silently making a consequential design decision.

Model identity raises operational questions

The conversation broadens into agent coordination, security boundaries, and the debate about AI consciousness. At 47:04, William brings the discussion back to model welfare and the language used in a model’s constitution.

We explore how those choices could affect the behavior of a system connected to production tools. William asks what an unexpected refusal would mean inside a network or cloud workflow. I acknowledge how difficult the consciousness question is and discuss my preference for tools grounded in knowledge and evidence.

These are our perspectives in a conversation about an unsettled subject. The engineering questions are concrete: what instructions shape the agent, what can it access, and how does the operator understand its decisions?

Smaller models, narrower decisions

At 1:00:04, we turn to TypeSafe’s System One models and Jev. I describe my experiments using decision scores and ranked outputs alongside NetClaw, and we discuss the possibility of testing candidate fixes in digital twins.

William also challenges the idea that a model cannot hallucinate. A numerical result still needs interpretation and validation. We finish this part of the discussion with open-weight models, local hardware, and the possibility of making useful AI more widely accessible.

Keep asking questions

Our closing exchange at 1:10:59 comes back to community: helping people get started, making room to experiment, and being willing to change your mind as you learn.

Thank you, William, for another thoughtful conversation, and thank you to Packet Pushers for sharing it. If you found this site through the episode, welcome. There is plenty here to watch, read, and build.

• Watch the full conversation on YouTube.
• Listen and read the official Packet Pushers episode notes.
• Explore NetClaw, MCP, and the VibeOps community.
• Find more Packet Pushers appearances and my GDE community work.

Published October 14, 2026. The episode is TCG086 on YouTube and Packet Pushers. All timestamps above refer to the 1:14:14 YouTube upload and were checked against its auto-generated transcript.