AI Agent Glossary: Platform Terms Defined

• Agent — a configured AI worker: one LLM, an explicit set of tools it may call, a memory profile, and a guardrail profile. The basic unit of execution. Not “any script that calls a model” — the configuration is the point.
• Tool — a connected integration (Salesforce, Slack, Postgres) exposing typed actions an agent can invoke. Tools are attached deliberately; an agent without a tool cannot reach the system, whatever its prompt says.
• Skill — a reusable, versioned prompt and instruction set. Attach one to many agents to share behavior without copy-pasting prompts that then drift apart.
• Knowledge base — indexed documents the agent retrieves from at runtime via RAG. Distinct from memory: knowledge is shared and curated, memory is per-agent accumulated context.

The full builder walkthrough is in the agents documentation.

• Workflow — an executable graph orchestrating one or more agents with triggers, conditions, loops, and approvals. Structure is declared up front, which is what distinguishes a workflow from an agent improvising a tool chain at runtime.
• Trigger — a workflow’s single entry point: webhook, schedule, chat, email, Slack, or manual.
• Condition — a deterministic branch node; routes execution based on a prior step’s output.
• MapReduce — the fan-out pattern: a mapper agent runs over each item of a list in parallel, a reducer collapses the results. For batch judgment at scale.
• Manager-worker — a manager agent decomposes an open-ended task and delegates to specialist workers. Trades some determinism for flexibility.
• Sub-workflow — a workflow nested as a node inside another; the parent waits for the child’s output.

Pattern selection guidance lives in the workflows documentation.

• Policy — a declarative, runtime-enforced rule about what agents can do: tool allow/deny, data access, rate limits, cost ceilings, time limits. Versioned, scoped, and immutable once active. Not a prompt instruction.
• Guardrail — a runtime inspector of content: prompts, tool arguments, tool responses, and final outputs, scanned in real time for PII, injection, toxicity, and topic drift.
• Plugin — one detection capability inside a guardrail profile, such as builtin.pii.scanner or builtin.prompt.injection. Each binding pairs a plugin with an action: block the run, regenerate the response, fix the content by redaction, or observe silently.
• Approval — a human-in-the-loop gate; execution pauses until a reviewer decides.
• Scope — the level a policy applies at: platform, tenant, or agent. Narrower scopes tighten, never loosen, broader ones.

The policy/guardrail distinction is the load-bearing one — see policies for how they compose.

• Run — one invocation of an agent or workflow, recorded with status, duration, and cost.
• Trace — the full execution record of a run: every LLM call, tool invocation, guardrail event, and decision, with timing and tokens. The unit of debugging.
• Eval — a measurement run scoring agent behavior. Four modes: simulation (synthetic scenarios), annotation (human-labeled outputs), A/B testing (two variants, one dataset), drift detection (continuous sampling of live traffic).
• Dataset — the inputs, with optional expected outputs, an eval scores against. Golden datasets are curated regression suites.
• Judge — whatever produces a metric: a model, a rule, or a human. Model judges get calibrated against human labels before being trusted.

Monitor records what happened; Eval scores whether it was right. Conflating the two is the most common vocabulary mistake in the field.

• Workspace — an isolated tenant with its own agents, workflows, tools, users, and billing. The multi-tenancy boundary.
• Role — a permission set. Built-ins are Owner, Admin, Member, Viewer; custom roles cover fine-grained cases.
• Channel — a delivery surface for an agent: web chat widget, Slack, email, CLI, or API. One agent can serve several.
• DAK — the Dezifi Agent Kit, a CDK-style infrastructure-as-code SDK (TypeScript, Java, Python). An App contains Stacks, stacks declare agents, workflows, and guardrail profiles, and dak deploy ships them. Everything the UI builds, DAK can declare.

If your team aligns on nothing else, align on agent, workflow, policy, guardrail, and trace — the five terms every production conversation runs through. The getting-started guide puts them in motion.