AI recruitment portfolios need a production-readiness story

An agent that demos cleanly can still be a liability in production. If you are an AI engineer interviewing for a delivery role, show one contained capability, the approval gate that stops it from writing business state, the CI baseline that catches a regression, and the confidence threshold that routes a decision to a human. The field guide maps those concerns as a maturity ladder, but it leaves the career problem open: how do you present that evidence without pretending a notebook demo is a shipped system?

What should replace the happy-path demo?

Show one capability as a fixed sequence of steps, with the model confined to a single node. The agent proposes; a separate component applies changes after approval. That sounds less impressive than an autonomous loop, and that is the point. If your current demo lets the agent write business state directly, stop there and refactor before recording a walkthrough. An interviewer can see the missing approval gate in a diagram faster than they can evaluate your prompt.
The same rule applies to loops. Bounded ReAct belongs in one place, not around every decision. If you cannot explain what ends the loop, you have not contained it.

Which check fails first when you add evaluation?

Evaluation fails first if behavior depends on the path the agent wandered. A CI gate that compares each release with a baseline is useful only after the graph is fixed. If you change a prompt and hope, the eval layer is still a notebook, no matter how many test cases it contains. Put the failing check in the portfolio: "prompt change regressed tool selection; baseline comparison blocked the release." That is stronger evidence than a screenshot of a green test run.
Shadow evaluations from production are a later step, not a substitute for a baseline. If you have no production traffic, say so and show the baseline you would use.

How do you present confidence and safety without production access?

Describe a composed, calibrated confidence score and the threshold that routes a decision to a human. If the agent always acts on a raw model score, the human-in-the-loop claim is false. The next step is abstention, not a better prompt.
Guardrails should be layered across input, output, PII, verification, and a judge, and they should fail closed. Sensitive data gets redacted or hashed at the boundary, and the audit ledger is append-only. If you only have one moderation filter, do not scale the agent; add the missing layers first. In an interview, a short failure table is more convincing than a feature list: guardrail block, override rate, token cost, and the person who can stop automated decisions without a deploy.

What should the walkthrough actually contain?

1. State the capability and the business constraint it serves.
2. Draw the fixed graph and mark the one model node.
3. Show the approval step that keeps the agent from writing state.
4. Show the CI baseline that catches a regression.
5. Name the confidence threshold, the human route, and the kill switch.

The guide's 90-second self-assessment is a useful reality check before you label a project production-ready: fewer than half the boxes ticked means you are earlier than the demo suggests. A portfolio that says "proposal-only, evaluated against a baseline, with a human route and an append-only audit ledger" is narrower than "autonomous agent platform," but it survives follow-up questions.
If you are going through AI recruitment, that distinction matters: a reviewer can check a proposal-only flow, a baseline, and an audit ledger; they cannot verify "autonomous" from a slide. Lead with the exact step that fails when the model is given more freedom, and explain what you removed to make the system dependable. Production readiness is not a bigger model or a longer prompt. It is the machinery that makes a small judgment call testable, and the judgment to show only the parts you can defend.