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

Claude Code sends a write through an MCP tool, receives a timeout, and verifies the external system before continuing, retrying, or stopping

A Claude Code tool timeout does not mean the action failed

A timed-out MCP tool call may have completed after the connection closed. Verify the external result before Claude Code retries a write.

An AI agent proposal passing through schema, operation, target, policy, dry-run, and approval checks before a receiving system may execute it

Treat AI agent output as untrusted input at every handoff

An approved AI agent can still produce a dangerous command, query, or tool argument. Validate generated output at the receiving boundary before another system acts on it.

An AI agent model fallback stopped at a security gate that checks provider, region, data, retention, tools, and policy before routing the request

Your agent's fallback model needs its own security decision

A model fallback can move prompts, retrieved data, and tool authority across a new provider or region. Treat each switch as a fresh policy decision, not a harmless retry.

An AI agent memory record passing through provenance, scope, expiry, review, and deletion checks before reuse

AI agent memory needs provenance and an expiry date

Persistent memory can make an AI agent faster, but stale facts, copied instructions, and missing ownership can quietly steer later runs. Store provenance, scope, review dates, and deletion rules with every memory.

An AI agent evaluation that checks the request, identity, tool path, data boundary, side effects, evidence, and final answer

Test the tool path before you trust the agent's answer

An AI agent can return the right answer after using the wrong identity, reading the wrong data, or calling a tool it never needed. Production evals should inspect the path as well as the result.

A Claude Code run replay record connecting the task contract, agent actions, decision points, evidence packet, human review, and incident replay

A diff cannot replay a Claude Code run

A Claude Code diff shows the final change, but it leaves out the starting state, failed commands, MCP results, skipped tests, and plan changes. Capture a compact replay record before the run disappears.

A production control loop for Claude Code with task contract, stop conditions, bounded run, evidence packet, and human review

Claude Code stop conditions are agent security controls

A production Claude Code run should know when to stop. Stop conditions turn unclear scope, wider tool access, weak rollback, and missing evidence into human review moments before an agent crosses a security boundary.

A four step control loop for normal agent grants, exception asks, human gates, and receipts

Give agent exceptions a budget before they become policy

Claude Code and enterprise AI agents sometimes need more access than the normal task contract allows. Treat that extra authority as an exception budget, with a reason, owner, expiry, evidence, and revocation path.

A Claude Code agent run leaving receipts for changed files, tests run, logs captured, risk notes, and undo path

Claude Code needs permission receipts, not blanket trust

Claude Code permissions matter before the run, but teams also need permission receipts after the run: what the agent used, what was blocked, what still needs approval, and how to undo it.

Two books connected to an operating stack for useful, reviewable, bounded, and auditable AI agent work

Treat agent autonomy like a production privilege

The safer question for Claude Code and enterprise agents is not how much autonomy the model can handle. It is which privilege the work deserves, how long it lasts, and what evidence the human gets back.