Improving AI agents from mistakes

TL:DR;

• AI agents need shared meanings when they work across different systems.
• An ontology defines business concepts, relationships, and rules.
• A semantic layer maps raw data to those concepts and defines trusted measures.
• Each failure can improve the ontology, measures, and tests, sometimes through another agent.

AI agents are moving into everyday business work

AI agents can search information, call tools, update records, and hand work to other people. Companies are beginning to use them for customer support, order management, finance, and maintenance.

As agents move into these workflows, their answers become part of how a company operates. An agent might decide which customer to contact, which order needs attention, or which machine should be inspected. These decisions depend on the information available to the agent and on what the company means by words like “late,” “urgent,” or “at risk.”

Different systems describe the same business in different ways

Most companies have many systems that contain pieces of the same story. An online shop stores orders, a delivery company stores shipping events, and a support system stores customer complaints. Each system uses its own fields and status names.

This creates problems when an agent has to connect the pieces. A delivery system might report that a package is held_at_depot. This means the package is stuck at a depot and may arrive late. The agent may understand that the package has been shipped, but miss the fact that it is unlikely to arrive on time. The answer can sound reasonable while leaving out a fact that changes the decision.

This is where antifragility becomes useful. Antifragility describes systems that benefit from shocks and surprises. Resilience means recovering after a problem. Antifragility means using the problem to improve the system. I explain the broader idea in Black Swans and AI.

Ontology and semantic layers give agents a shared view

An ontology is a map of the things a business cares about and how they relate to each other. For an online shop, it might describe customers, orders, products, deliveries, and complaints. It can state that an order belongs to one customer and that a delivery belongs to an order. It can also describe a delivery hold and connect it to the affected delivery and order.

The ontology provides categories, relationships, and rules for the company’s raw order records. Those records are mapped to that structure and checked against it. If a new record or status has no place in the ontology, it should remain unknown until someone gives it a meaning.

The semantic layer connects raw data to this structure. Raw data consists of database rows, event messages, files, and other records produced by daily operations. The semantic layer maps those records to business concepts and defines shared measures, such as “late order” or “orders needing attention.” When an agent asks for one of these measures, the semantic layer runs the agreed query over the underlying records.

For example, the ontology answers:

• What is an order?
• What is a delivery?
• How are orders related to customers in general?

The semantic layer answers:

• When do we call an order late?
• When does a delivery need attention?
• How should we count late orders by customer?

The agent uses the two layers for different jobs:

sequenceDiagram
    participant U as User
    participant A as AI agent
    participant O as Ontology
    participant S as Semantic layer
    participant D as Raw data

    U->>A: Which orders need attention today?
    A->>O: What are orders, deliveries, and delivery holds?
    O-->>A: Concepts, relationships, and rules
    A->>S: Run the trusted “orders needing attention” measure
    S->>D: Query order and delivery records
    D-->>S: Raw operational facts
    S-->>A: Matching orders, definition, and evidence
    A-->>U: Explain the result with evidence

The agent asks the semantic layer to apply the shared definition and return the matching records. The semantic layer keeps the meaning of the measure in one place and provides individual records as evidence. The model may change, the source systems may use different names, and the agent may choose different steps to solve a task. The meaning of the task still comes from the same shared definitions.

Operational failures can be used to improve the shared view

Imagine an agent answering the question, “Which orders need attention today?” It looks at the order system and the delivery system. The semantic layer defines an order as needing attention when its promised date has passed without a delivery confirmation.

One day, a delivery company introduces the status held_at_depot. The status reaches the raw data store, but nobody has connected it to the business meaning of a delayed delivery. The agent sees that the package has left the warehouse and does not flag the order. The customer complains the next day.

The company can now improve the system in a lasting way:

• The ontology adds a DeliveryHold concept and connects it to the delivery and the order.
• The semantic layer maps held_at_depot to DeliveryHold.
• The definition of “needs attention” includes an active delivery hold.
• The original question becomes a test that the system must answer correctly in the future.

The next time the agent sees the same status, it flags the order and explains why. If it encounters an unknown status, it can send the case to a person instead of silently guessing. The company learned from one operational mistake and improved the shared map, the business measure, and the check around them. That is what antifragility looks like in an AI system.

These updates can be made by another agent. An improvement agent can inspect wrong answers and human corrections, spot the unmapped status, suggest an ontology change, update the mapping and measure, and create a new test. A person can approve the change, or a controlled system can apply low-risk changes automatically. This opens a path toward self-learning systems. The system learns by improving its shared concepts, metrics, and tests from experience, even when the underlying language model stays the same.

As agents move into the core of company operations, this shared view becomes increasingly valuable. Agents will encounter new data, changing processes, model updates, and unexpected cases. Companies should make sure that each important failure can improve the definitions that guide future decisions. This lets agents change over time while the business retains a clear understanding of what their decisions mean.