A system that can answer, reason, plan, use tools, and take action — not just chat.

AI that acts on real systems, not just answers questions.

Policies that decide what an agent is allowed to do, not just what it can see.

Policies that decide what data an identity is allowed to see.

A structured record of what the agent did, why, and with what evidence.

A shared inbox where humans review agent actions before they execute.

The structured list of every action an agent can take, with risk tier and owner.

An in-line service every agent action passes through for policy checks and logging.

A multi-agent topology where agents call each other freely. Powerful and easy to overuse.

The layer that captures how your company defines its world, beyond just schemas.

A condition or exception the business enforces, like discount eligibility.

Slowing the producer when the consumer cannot keep up.

How many people, records, or systems an action affects if it goes wrong.

Reusing a previous answer to avoid recomputing it.

An inventory of where data lives and who owns it.

Trying a cheap model first and escalating to a larger one only when needed.

A reference the agent shows so users can verify where an answer came from.

The ability to swap models, tools, or vendors without rewriting your system.

How sure the agent is about its answer, ideally surfaced honestly.

The meaning, definitions, and relationships an agent needs to answer correctly.

A data professional who owns meaning the way engineers own code.

The queryable layer above storage and compute that holds business meaning.

The maximum amount of text a model can read in a single request.

The set of policies that make agent behavior safe and predictable.

The total expense of running AI — tokens, tools, retries, and infrastructure.

A design that treats cost like latency: a first-class requirement.

Small, reliable tools that agents combine to do bigger work.

Confidence derived from system evidence, not from raw model probabilities.

A piece of data offered with quality, ownership, and a clear contract.

An open table format that adds reliability features on top of Parquet files.

When model or data behavior changes over time, often silently.

Same input, same output, every time. Hard for LLMs, important for audits.

A policy outcome that lets a smaller, safer version of the action proceed.

A numeric representation of text used for similarity search.

Measuring whether the agent is giving correct, safe, useful answers.

A test suite that scores AI quality across many examples.

Building AI features by writing evaluations first, then improving the system.

Adapting a base model to your domain by training it on your data.

A single entry point that all AI calls go through for security, routing, and logging.

A list of business terms with agreed, simple definitions.

The set of rules that decide what is allowed in the data and AI stack.

Anchoring an answer in real, citable data instead of model memory.

Code-enforced rules that block dangerous inputs or outputs.

A curated set of examples used as the truth for evaluation.

When a model produces something that sounds confident but is not true.

A design where humans review, approve, or take over agent actions.

The typed schema two agents agree on when passing work between them.

An open table format widely adopted for lakehouses.

Bringing raw data into your platform from source systems.

Asking a model to produce an output, as opposed to training it.

Safe to retry — the same call twice gives the same result.

The compounding velocity gain a platform earns by routing change through stable interfaces.

A structured map of concepts and how they relate, used to enrich context.

A storage architecture that combines lake flexibility with warehouse reliability.

The path a piece of data took from source to answer.

A large language model trained to generate human-like text.

Model Context Protocol — a standard way for agents to talk to tools and data.

A defined business number, like revenue or churn.

The place where metrics are defined once and calculated consistently.

Sending each request to the smallest model that can handle it well.

A map of tasks to the right model size for each one.

A design where multiple specialized agents work together.

The ability to see what your AI system is doing in production.

A structured map of business concepts and how they relate.

A data or model format anyone can read without vendor lock-in.

A standard API anyone can implement, reducing lock-in.

Coordinating multiple steps, tools, or agents into one workflow.

An open columnar file format widely used in data platforms.

A rule about what an identity is allowed to see or do.

Writing governance rules as software the system enforces automatically.

A service that evaluates policy-as-code at runtime.

The instructions and context sent to a model in a single request.

Designing prompts so models behave well for a given task.

A multi-agent topology where agents run in a fixed sequence — the safest default.

The chain of sources behind an agent's answer, surfaced so users can verify it.

Retrieval-Augmented Generation — letting the model read your data before answering.

A model thinking through steps before producing a final answer.

A shared pattern teams follow so each new project does not reinvent the wheel.

Looking up relevant content to give a model better context.

Re-running a failed model call. Quietly expensive at scale.

Turning off or reverting an agent quickly when something goes wrong.

A cheap classifier that decides which model or tool handles a request.

A system where you can re-run history to debug or recover.

Whether an action can be undone cleanly, and how quickly.

The shape of a table — columns and types.

A cache keyed by meaning, not exact text, that reuses similar answers.

The layer that maps business meaning to underlying data.

Finding content by meaning, not exact keywords.

A label that marks data by how sensitive it is, like PII or financial.

A measurable indicator used to monitor quality, cost, or safety.

The one place a fact is considered authoritative.

The ongoing care and ownership of a data asset.

Continuous data that arrives event by event rather than batch by batch.

Looking up answers in tables, metrics, or APIs before falling back to text search.

Generated data used for training or testing when real data is scarce.

Something an action changes in the outside world, like sending an email.

A multi-agent topology where one agent routes work to specialists.

Data emitted by a system that lets you observe its behavior.

Limiting the rate of requests to control cost or load.

The unit a model reads or writes — roughly a few characters.

A target for how many tokens a request is allowed to use.

When the agent invokes an external tool, like a query or API.

The full record of a single agent run, end to end.

The sequence of checks an agent action passes before it executes.

A nine-step reference flow every production agent follows.

A dataset agreed upon as authoritative for a given metric.

The explicit rule that ends an agent loop. Without one, loops burn money.

A visible cue (source, definition, confidence, limit) that helps users decide when to trust an agent.

Cost per unit of value delivered, like cost per resolved ticket.

Databricks' governance layer; one example of a managed catalog.

Checking that an input or output meets the rules before using it.

A store that lets you search by embedding similarity.

Finding similar content using embeddings.

A situation where switching providers is expensive or slow.

Tracking changes to data, models, or definitions over time.

A structured store optimized for analytics queries.

A defined sequence of steps a system runs to complete a task.

Asking a model to do a task it was not explicitly trained for.