# Glossary These terms describe Operade's current text-agent model. They are operational definitions, not promises about future product scope. ## Tenancy and agent configuration ### Workspace The tenant boundary for people, roles, agents, knowledge, tools, usage, and credentials. API keys are bound to one Workspace, and resources outside that scope appear not found. ### Agent A named text-based assistant whose authored configuration can include prompts, model settings, Knowledge, Tools, Procedures, Workflows, Guardrails, and Analysis. An Agent is the continuing identity; its Versions are immutable snapshots. ### Draft Mutable, unpublished Agent configuration for one user on one Branch. A Draft remembers the base Version from which editing started. If the Branch advances, the Draft may need reconciliation before publish. ### Branch A named line of Agent history with a current tip Version and an optional share of deployment traffic. Main is the primary Branch. Protection, archive state, role, and live traffic constrain which mutations are allowed. ### Version An immutable snapshot created by publishing a Draft or by a history operation that produces a new snapshot. Existing Conversations stay pinned to their selected Version or frozen draft context. ## Runtime and evidence ### Deployment An append-only record of the complete traffic distribution across Branches for an Agent. A valid distribution totals exactly 100 percent; changing or rolling back traffic creates a new Deployment record. ### Conversation A sequence of text turns for one Agent, pinned to a Branch and Version or to a frozen Test Chat draft. It has a source, status, timestamps, and an end reason when closed. ### Event An ordered, append-only runtime evidence item attached to a Conversation. Events can describe model requests, Tool calls, Workflow transitions, Procedure state, Knowledge retrieval, and Guardrail outcomes while keeping unsafe values out of public payloads. ### Knowledge Workspace content that can be attached to an Agent and retrieved as data during a Conversation. Knowledge is indexed from supported text, files, or public URLs and is not treated as trusted system instruction. ### Tool A callable capability, currently including governed webhook requests and runtime actions. Tools define parameters, optional authentication through named Secrets, and safe output handling. ### Test Chat An interactive text conversation used to explore one saved Draft or published branch tip. It is not a reusable evaluation and does not produce a batch pass rate. ### Agent Test A workspace evaluation definition of type Simulation, Next Reply, or Tool Invocation. Edits create immutable revisions so a past result stays tied to the definition that produced it. ### Test Invocation A durable batch request that freezes an explicit Main Draft or Live Agent configuration, selected test revisions, Tool snapshots, and repeat count. ### Test Run One execution of one invocation item at one repeat index. A run records status, result evidence, usage, latency, and an optional failure code. ### Failure bucket A post-run grouping of similar terminal failures across repeated runs. Bucketing supplements the raw result; it never replaces or rewrites that evidence. ## Authored control flow and evaluation ### Procedure Reusable customer-service guidance authored as free-form instruction or structured steps. A Procedure can ask, tell, say, call a Tool, branch, or invoke another Procedure within tested limits. ### Workflow A graph of nodes and ordered conditional edges that controls agent execution. The first matching outgoing edge wins, and runtime limits prevent unbounded transitions or repeated subagent execution. ### Guardrail A configured input or output safety evaluation. Guardrails can allow, block, or retry according to the configured action; evaluator failures stop safely instead of exposing unreviewed content. ### Analysis Post-conversation structured evaluation that can produce a summary, result, and rationale. Analysis has explicit states such as pending, processing, done, skipped, or failed and contributes cost only when it reaches a model request. For the lifecycle connecting these terms, see [Core concepts](/docs/agents/core-concepts), [Agent Testing](/docs/agent-testing), [Deploy overview](/docs/deploy/overview-and-branches), and [Transcript and Activity](/docs/monitor/transcript-and-activity).