Sources

    Who's Watching the Agents?

    AGENT UX · CASE STORY

    Who's watching the agents?

    What it takes to let AI agents act inside your apps without losing control of them.

    4 min

    Imagine giving a new hire the keys to your office.

    You wouldn't hand over every key. You'd decide:

    Which doors they can open

    When they need a manager's OK

    Who signs the log

    How to take the keys back

    AI agents get keys too.

    Usually with none of those rules.

    Meet Priya. She runs five apps, and twenty-four AI agents work inside them.

    A chatbot answering visitors

    A categorizer sorting bank transactions

    A writing assistant

    A spend tracker

    +20 more

    Priya

    Who's watching all of them?

    Watching isn't the same as deciding.

    WATCHING

    What did it do?

    Dashboards tell you after the fact.

    DECIDING

    What may it do?

    Rules apply before it acts.

    Priya needs both. Here's how Warden does it, one rule at a time.

    Rule 1

    Every agent gets its own keys.

    Agent

    Finance app

    Categorizer

    Trust: Supervised

    CAN

    Read transactions

    Write categories

    CAN'T

    Move money

    Touch bank links

    PRINCIPLE

    Least privilege

    Give each actor only the access its job needs, and deny everything else by default.1 An agent is not its owner. It gets its own, smaller identity.

    Rule 2

    Every action passes five gates, in order.

    ✓

    Is the agent switched on?

    ✓

    Does its owner have the authority?

    ✓

    Is this inside its scope?

    !

    Does a guardrail say stop?

    ✓

    How much can it do on its own?

    Allowed

    Needs approval

    Blocked

    Autonomy is checked last. However much an agent is trusted, it can never go beyond its keys.

    Rule 3

    Guardrails beat trust.

    Production needs approval

    Limit how much it can touch

    Rate limit

    Ask when unsure

    Anything that speaks for her waits for her yes

    The strictest rule wins. A guardrail can stop even a fully trusted agent.

    Rule 4

    When an agent stops to ask, ask the right way.

    Held action

    Ask when unsure

    Categorizer · invoke categorize · staging

    Approve

    Deny

    Priya

    What exactly am I approving?

    Agents ask two kinds of questions. They need two kinds of answers.

    PERMISSION

    "May I re-sort 9,000 transactions in production?"

    Approve or deny.

    KNOWLEDGE

    "What is this $500?"

    Give it the answer.

    TRANSACTION

    ZELLE TO P RAO

    $500.00 · Aug 12 · checking

    Family transfer

    Rent

    Gift

    Other

    ☑ Make this a rule, so it never asks again

    Resolve

    PRINCIPLE

    Ask the right way when unsure

    When an AI is uncertain, it should ask the person instead of guessing.2 The control should match the question, and the card should show the thing being acted on, not just the verb.

    Rule 5

    An off switch that really works.

    Agent

    Live

    Website chatbot

    Answering visitors on the live site.

    Pause agent

    Pause it and the live chatbot stops answering within seconds. The page stays up. Only the AI stops.

    PRINCIPLE

    Easy to stop, easy to undo

    People need a clearly marked emergency exit.3 For agents, that means a pause that takes effect in the real app, not just on a dashboard.

    Rule 6

    Every decision answers: on whose authority?

    Audit

    Re-sort 9,000 transactions

    ✓

    Agent switched on

    ✓

    Owner: Priya

    ✓

    In scope

    !

    Guardrail: production needs approval

    Held for Priya

    Every attempt is on record with the reasoning that allowed or stopped it.

    And it watches

    What the agents cost, and who they serve.

    AI spend per agent

    Visitors, with no third-party trackers

    Cost per visitor

    Uptime for every app

    Watching and deciding, in one place.

    Priya

    I know what every agent can do, what each one did, and I can stop any of them.

    BEHIND THE STORY

    Priya's setup is mine.

    Warden is a personal tool I built. It governs twenty-four AI agents across five of my apps. The chatbot on my portfolio asks Warden for permission before every free-form answer.

    I started it to test the waters for a potential Datadog interview. Then the problem got more interesting than the interview, and I couldn't stop.

    And the Resolve control? It started with me looking at a held transaction and asking: what exactly am I approving?

    CHEAT SHEET

    Letting AI agents act, safely

    Own keys per agent. Deny by default.

    Check in order: on, owner, scope, guardrails, then autonomy.

    Guardrails beat trust. The strictest rule wins.

    Two kinds of questions: yes/no for permission, an answer for knowledge.

    A real off switch.

    Record who decided, and why.

    TRY IT

    Explore Warden.

    The public demo shows the whole console with clearly labeled sample data.

    EXPLORE WARDEN →

    You finished case story 01: "Who's watching the agents?" Warden is a personal tool by Anuja Nimmagadda, built by directing Claude Code. Priya and the examples in this story are fictional.

    Sources

    1. Saltzer & Schroeder (1975). The Protection of Information in Computer Systems. Proceedings of the IEEE 63(9).
    2. Amershi et al. (2019). Guidelines for Human-AI Interaction. CHI 2019 (Microsoft HAX).
    3. Nielsen Norman Group. 10 Usability Heuristics for User Interface Design.