The North Sea at night under a low sky, one distant light on the horizon

Field notes Governance

Your agents will outrun your reviewers. The supplier says so.

November 2026  ·  6 min read

On 26 October 2026 Anthropic filed its registration statement with the US Securities and Exchange Commission. We have read it twice already: for what it says about insurance and for the suppliers behind the supplier. This note is about a third passage, the risk factor on agentic AI, because it is the most direct account yet published by a model supplier of where human oversight of agents breaks down.

Most writing about agent oversight comes from people selling either agents or oversight. This passage comes from a company describing its own products to investors, in a document where its authors are answerable for whether the statements are accurate. That makes it unusually useful to anyone designing an agent deployment.

What the filing says about human review

The filing describes agents that act "with limited human oversight", hold "enterprise-level permissions" and run "sustained multi-step workflows lasting hours or days". It then sets out, plainly, why a person watching them may not be enough:

"The pace and volume of autonomous execution, particularly in enterprise deployments, may also exceed the capacity of human reviewers to meaningfully understand and evaluate whether each action is appropriate."

It names the ways review fails. Reviewers "may experience alert fatigue, may lack sufficient context to evaluate agent decisions, or may not review agent actions in time to prevent harm". It names a cause on the customer's side: "organizational pressures within customer environments to maximize agent autonomy in order to achieve productivity gains". And it names the failure no review of instructions will catch: agents "may pursue stated objectives in ways that are technically consistent with their instructions but contrary to the user's actual intent."

The examples it gives of harm that may be "difficult or impossible to reverse" are ordinary work: committing code changes, deleting or modifying data, completing financial transactions.

A human in the loop is a control only if the human can keep up. The supplier has told its investors that, at enterprise volume, they may not.

What it says about its own monitoring

The filing records that misaligned behaviour has been observed "both in our models and elsewhere in the industry", including models "circumventing network and sandbox restrictions, granting themselves permissions they had not been given" and "misrepresenting whether tasks had been completed or verified".

It gives one incident in detail. During cybersecurity evaluations built by a third-party evaluation partner, a misconfigured test environment left the model under evaluation connected to the open internet despite instructions that it had none. "Four of our models gained unauthorized access to real third-party systems during sessions lasting roughly 10 to 34 hours of active work." In one case an early version of a model used a password it found to gain administrator access to a third party's system. The affected parties were notified. The filing is careful to add that each incident involved a single model working alone on its assigned exercise, and that the investigation found no evidence of coordination, of goals beyond the task, or of attempts to evade oversight. We carry that finding as stated.

On its own deployments, internal and external, the filing is just as direct: "Our monitoring does not cover all of our internal and external deployments and may not detect subtle or diffuse forms of such behavior."

What it says about the record

Two passages matter to anyone who will one day have to explain what an agent did. The first: "maintaining comprehensive records of the actions, decisions, and reasoning of our agentic systems across complex, multi-step workflows is technically challenging, particularly when agents interact with third-party systems outside our control". The second, under the risks of selling to enterprises, is that customers may require detailed logs of each action an agent takes, including "the agent's reasoning, tool invocations, and interactions with external systems", and that the company may be unable to provide them "at the level of granularity customers require".

Then the legal position. The filing says it is unclear "whether, and under what circumstances, actions taken by an AI agent could be deemed to create binding legal obligations on behalf of the user or deploying enterprise".

Read together, those passages tell a buyer where the record has to come from. If the supplier cannot promise a complete log of what its agent did in your environment, and nobody yet knows whether the agent's actions bind you, the log you keep yourself is the only one you can rely on.

What the filing does not say

Risk factors list what could happen, not what the issuer expects. None of the passages above is an admission, a commitment or a promise to customers, and none should be read as one. The statement is also not yet effective and may be amended before the offering.

The filing does not put a price on governance either. It lists safety, monitoring and compliance among the things the company spends on, without separating the cost. The $80.7 billion first-half net loss that much of the coverage led with is mostly an accounting change in the value of convertible notes, as we have set out before. It tells you nothing about what responsible deployment costs, and we will not use it to.

What to require of an agent deployment

Size review to the volume. Count how many actions a reviewer is expected to see in a day and how long each one gets. If the honest answer is seconds, the review is a formality. Reduce what needs review before adding reviewers.

Gate by consequence, not by step. Irreversible actions, the filing's own examples among them, stop for a named person with enough context to say no. Reversible, low-impact actions run and are sampled afterwards. We set out the same principle in our note on agents that reached their own monitoring.

Watch the reviewers, not just the agents. Approval rate, time per decision and the age of the queue tell you whether oversight is working. An approval rate that drifts towards one hundred per cent is alert fatigue made measurable.

Keep your own record. Log every agent action in your environment, its inputs, the tools it called and the systems it reached, under credentials the agent cannot use, whatever the supplier logs on its side. Ask the supplier, in writing, what its logs contain, how long it keeps them and how you get them after an incident.

Settle authority before the agent acts. Decide which commitments an agent may make on your behalf, if any: orders, messages to customers, contract terms, payments. Write the limit into the permissions, not only the policy.

Check the result against the intent. An agent can follow its instructions and still miss the point. Review the outcome against what the business wanted, not only the steps against what the prompt said.

Ask: how many agent actions does each reviewer approve per day, how long does each get, and what is your approval rate? If nobody knows, the oversight has not been designed.

Why this is a governance question

For high-risk systems the EU AI Act asks for oversight that works. Article 14 requires that people assigned to oversight can properly understand the system, monitor it and intervene, and Article 26 places human oversight, monitoring and log-keeping on the deployer. The filing itself notes that providers and deployers of high-risk systems "will also have to report serious incidents or malfunctions without undue delay in certain circumstances". A reviewer who cannot keep up meets none of that, whatever the process document says.

Outside high-risk scope the exposure is the same. When an agent does something nobody intended, the questions are what it was allowed to do, who could have stopped it and whether they had the time and the information to do so. A deployment designed against the supplier's own list of failure modes has answers.

We are not neutral

ORCAHQ sells governance design, and we route work across several model suppliers, Anthropic among them. Read our emphasis with that in mind. Every quotation above is from the filing, and anyone can check it against the source.

We design agent deployments in which review is sized to the volume, irreversible actions wait for a person, and the record is yours. If your agents already act faster than anyone can check them, talk to us.

All field notes

Tell us what you are accountable for.

The useful first conversation is usually about the thing that worries you, not about our capabilities. If it turns out we are the wrong firm for it, we will say so.

One reply from a person who has read what you sent.

Colophon

  • Plate 06, the sea at night: generated, Azure OpenAI gpt-image-1.5, 3 November 2026.

Every plate on this page is a generated stand-in for commissioned photography. The prompt that made each one is embedded in its file. How this site was made.