Glossary

Terms, in plain words.

Agent: A software system that takes a goal, decides on steps, and uses tools to carry them out.
Agent run: One complete attempt by an agent at a goal, from the first input to the final output.
Attribution (cost): Assigning spend to the step or feature that caused it.
Audit trail: A tamper-evident record of what a system accessed and did.
Benchmark saturation: When top systems score so highly on a test that it no longer distinguishes between them.
Context: The information an assistant has available while working on a task.
Determinism: The property that the same inputs produce the same output every time.
Evaluation: A structured way of measuring how well a system does a defined piece of work.
Explainability: The ability to show why a system produced a particular result.
Fine-tuning: Further training a model on a narrower set of examples.
Frontier (of a run): The set of recorded values entering a run from outside it, such as model outputs and tool results.
Guardrail: A rule or check that limits what a system can do.
Hallucination: Output that sounds plausible but is not supported by the source material.
Local-first: Designed so the core work happens on the user’s own device.
Prompt injection: Instructions hidden in content an assistant reads, intended to change what it does.
Provenance: The recorded origin and history of a piece of data or content.
Red teaming: Deliberately trying to make a system fail, in order to find weaknesses.
Replay: Re-executing a recorded run against its own stored records to reproduce it.
Retrieval: Looking up relevant material at the time a question is asked.
Scoped access: Permission limited to what a particular task needs, for as long as it needs it.