Who Checks the Agent? Why Oversight Is Now the Hard Part of AI Automation

# Who Checks the Agent? Why Oversight Is Now the Hard Part of AI Automation

Building an AI agent has become the easy part. Knowing what it did last Tuesday, and whether it should have done it, is much harder. Two announcements in the past fortnight show how seriously the market now takes that gap. On 29 October, Crux Labs came out of stealth with more than US$6 million to turn expert judgment into checks on enterprise agents. On 3 November, CrowdStrike agreed to acquire Fiddler AI, an agent observability company, in a deal reportedly worth about US$400 million. One is a startup just out of stealth and the other a global security vendor, but both are betting on the same question: who checks the agent?

The 1% Problem in AI Agent Oversight

AI agent oversight is the set of controls that decide what an agent may do, who must approve its consequential actions, and how anyone can later prove what happened. Most teams are still doing it by sampling, and sampling leaves most of the work unseen.

SiliconANGLE's coverage of Crux Labs puts the problem plainly: enterprise agents can handle anywhere from thousands to millions of interactions a day, and many teams end up sampling the top 1% or less of traces for manual review. The vast majority of agent failures pass by unnoticed. At a million interactions a day, a 1% sample still means ten thousand traces for someone to read, and 990,000 that nobody opens.

Reviewing more traces by hand will not fix this. Agent volume grows faster than any team of reviewers can, so the answer has to come from how the work is designed and how it runs.

Why Valid Access Is Not the Same as Correct Behaviour

Traditional security asks whether a user or system is allowed in. That question is necessary for agents, but it is no longer enough. In its announcement, CrowdStrike notes that an agent can hold valid credentials and approved tools and still make unsupported decisions, ignore instructions, or claim work it never completed. It does not have to be compromised to create risk.

That changes who needs answers. It is no longer only the security team. "When a regulator or your board asks what an agent did, saying 'the model did it' is not an answer," said George Kurtz, CEO and founder of CrowdStrike.

Crux Labs approaches the same gap from the quality side. "Agents are incredible at attaining goals, but they lack access to the years of heuristics that we as humans have collected in our heads," said Julie Kheyfets, Crux Labs cofounder and CEO. An agent can follow every rule and still produce work your best people would reject.

The Four Checks Every Agent Needs

Oversight is often treated as one product category, but it is really four separate jobs that happen at different moments. A gap in any one of them leaves a question nobody can answer.

Observability and evaluation tools mostly work on the last two rows. They are valuable, but they generally describe or score work that has already happened. If an agent has already released a payment, a good trace explains the mistake without undoing it. The first two rows are what stop the mistake from going out.

Where Human Approval Actually Belongs

Human in the loop AI works when people review the few decisions that carry real consequences. Ask someone to approve every small step and the gate turns into a reflex: they click through, and the control exists only on paper.

A simple way to place approval gates is to sort actions by what a mistake would cost:

1. Reversible and low impact. Drafting a summary or sorting a document. Let the agent run, log the step, and review samples.
2. Reversible but visible. Updating a record or posting an internal notice. Let the agent run within set limits, and flag anything outside them.
3. Irreversible or high impact. Releasing funds, sending something to a client or regulator, or changing a contract. Stop and wait for a named person.

Take a progress claim on a construction project. The agent can match the invoice to the purchase order, verify the vendor's bank details and check the budget without anyone watching each step. Releasing the money is the moment a person should approve, with those completed checks in front of them. Reviewers spend their time on the one decision that matters, and they can see the reasoning behind it.

Oversight Has to Be Built In, Not Added Later

You cannot audit a control that never ran. If an action was not checked before it executed, no amount of later logging turns it into a governed decision. It only gives you a better description of an ungoverned one.

This is why Agento treats oversight as part of how a workflow runs rather than a dashboard added on top:

• Human approval gates. Sensitive steps wait for a person's sign off before they proceed, so high consequence actions never run on their own.
• Evidence capture. Each step records its inputs, the policy decision that governed it, and its outputs, so a reviewer can follow what happened without interviewing anyone.
• A full audit trail. Every action in every run goes to a tamper evident record, giving compliance teams something to verify instead of something to reconstruct.
• Rules that apply everywhere. Guardrails are set once and enforced on each action, with least privilege as the default, so an agent gets only the access its task needs.

The audit trail also makes evaluation more useful. Any quality check or outside monitoring tool is only as good as the records behind it, and a run that was captured properly can be reviewed, scored and explained long after it finished.

Five Questions to Test Your Oversight

Before giving an agent more autonomy, try answering these for one live workflow:

1. Could you pull the full record of a specific run from last month, including what the agent saw and did?
2. Which actions in this workflow cannot be undone, and does a named person approve each one?
3. Are permissions enforced before each action, or only reviewed afterwards?
4. What share of runs does anyone or anything actually check, and how would you notice quality drifting?
5. If a regulator asked why the agent made a decision, could someone who was not there answer from the record alone?

If several answers are "not sure", the agent has more autonomy than your oversight can support. Fix that before you scale it.

Frequently asked questions

Take the next step

Agents will keep getting more capable, so the teams that scale them safely will be the ones who can show what every agent did and why. Agento runs your business processes across the tools you already use, with a person approving anything sensitive and every action recorded and auditable.

Explore Agento → www.agento.au