Mcp | Thomas Talks AI: Claude Architecture, AI Agent Security, Claude Code & Production Engineering

A production AI agent flow with request, approval, action, evidence, and revocation path

If the agent can act, it must be easy to stop

Production AI agents need a revocation path before they get wider authority. If a Claude Code run, MCP tool, workflow agent, or RAG assistant goes wrong, the team should know exactly how to stop it.

An agent permission change moving through proposal, review, tests, approval, rollout, and rollback

Permission changes are releases

Every new agent permission is a production change. Treat new MCP methods, data sources, write paths, credentials, and approval bypasses like releases with evidence and rollback.

A bridge from a Claude Code run to an enterprise AI agent, connected by one control record

The control record is the bridge

Claude Code rollout and enterprise AI agent security need the same artifact: a control record that explains scope, authority, evidence, approval, and rollback.

A production AI agent rollout flow showing scope, authority, evidence, and release controls before autonomy expands

Put agent security before the rollout

Production AI agents need scoped authority, run evidence, approval gates, and rollback before the rollout expands. This is where Claude Code delivery work and enterprise agent security meet.

A production AI agent evidence loop showing task contract, bounded tools, run record, approval gate, and rollback path

If the agent cannot explain the run, do not scale it

Production AI agents need more than good output. Before a team scales Claude Code or enterprise agents, the run must explain its task, tools, evidence, approval, and rollback path.

A two-book production and security loop diagram showing Claude Code delivery controls and enterprise AI agent risk controls

No approval, no autonomy

If an AI agent can read internal data, call tools, change code, open tickets, or trigger workflows, it has delegated authority. Give it an approval packet before it acts.

A diagram showing a production agent surrounded by task scope, identity, tool boundaries, evals, observability, approvals, and rollback

The agent control plane is the product

A production AI agent is only as good as the control plane around it: scope, identity, tool boundaries, evals, observability, approvals, and rollback. Better prompts do not replace that operating model.

A left-to-right flow showing task and scope, shadow mode where the agent plans but writes nothing, human review of the plan, and execution where writes happen

Run the agent in shadow mode before it changes anything

The agent’s diff is the most expensive way to learn what it was planning to do. Run it once in shadow mode, where it produces a complete plan but executes nothing, and review the plan before you let it touch anything.

A left-to-right flow showing approved scope, scope expansion, stop and request, human approval, and a run record entry

A new permission means a new approval

A Claude Code run that widens scope after a one-time approval is working under an authorization you never gave. A new permission should mean a stop, a request, and a logged human approval, not a footnote in the summary.

A diagram showing an AI agent connecting through an MCP server to business systems, with scope, credentials, approvals, logs, and stop rules as controls

Your MCP server is part of the security boundary

MCP servers are not harmless connectors once agents use them to reach tickets, data, APIs, deployment tools, or RAG systems. Treat them as part of the security boundary.