Glossary

A quick reference for the vocabulary in these docs and the product.

Agent

One definition of an AI worker: its model, prompts, parameters, tools, knowledge, and settings, with a revision history. Edited as a form in Studio. The unit of work in Fruxon. See Agents.

Token

An organization-scoped, scoped access token for server-to-server access — a personal access token (fx_pat_) or a service-account token (fx_sat_). Sent via the Authorization: Bearer header. See Tokens.

Asset

A file or data source attached to an agent for retrieval or templating. PDFs, DOCX, Confluence pages, GitHub repos, web URLs. See Knowledge Base.

Audit Log

An Enterprise, workspace-wide record of who did what: every change made through the API, reads of personal data, and every time Fruxon staff access the workspace. Each event records the actor, action, resource, outcome, IP address, and time. Owners and Admins can filter and export it; it's kept for two years. Credential access has its own trail, the credential audit (see Security). See Audit Log.

Baseline (Evaluation)

The currently deployed revision, which a candidate revision is automatically compared against during an evaluation run.

Budget

A per-agent spend limit on model cost, for a day, a week, or a month, with alert thresholds. It either blocks new runs once reached or only notifies. Called a spend limit in the UI. See Cost, Budgets & Credits.

Candidate (Evaluation)

The revision being tested in an evaluation run, scored against the baseline.

Participant

A person, group, or AI agent that can address agents through one or more channels. See Agent Network.

Code Executor

A built-in tool that runs Python in a secure sandbox. Useful for transforms, charts, calculations.

Agent Network

Tenant-level wiring for triggers, participants, participant channels, conversations, and access requests around agents. See Agent Network.

Conversation

An observed provider thread or end-user session, such as a Slack channel, Telegram chat, Teams thread, or phone-number conversation. See Conversations.

Dataset (Evaluation)

A list of test cases — inputs with optional expected outputs and metadata — used to score agent revisions. See Evaluations.

Deployed Revision

The revision currently serving production traffic for an agent. Switch instantly by deploying a different revision. See Versioning.

Execution / Run

One execution of an agent. Produces a trace.

Execution History

The log of all runs for an agent, filterable and searchable. See Observability.

Firebase JWT

The web app's user authentication token. Carries the organization claim. Used for human users; tokens are used for server-to-server.

Golden Dataset

A curated dataset of test cases used as a quality benchmark for evaluations.

Integration

A configured connection to an external service (Salesforce, PostgreSQL, etc.). Exposes one or more tools to agents. See Integrations.

Judge

An LLM that scores evaluation outputs against your defined metrics.

Knowledge Base

The collection of assets and data sources attached to an agent for retrieval. See Assets.

LLM Provider

A vendor of large language models — OpenAI, Anthropic, Google, etc. Connect with your own API key. See AI Providers.

Memory

Per-agent persistent notes — each a title, body, and optional tags — searched by keyword. Read and written by the agent through tools. Survives across conversations. See Memory.

Metric (Evaluation)

A measure used to score outputs in an evaluation — correctness, faithfulness, format, tone, etc. See Evaluations.

Access Request

A proposed participant or participant-agent binding raised when an unknown or unbound sender tries to address an agent. See Access Requests.

OpenAPI Tool

An integration tool defined declaratively from an OpenAPI/Swagger spec — URL, method, parameters, auth.

Organization

The top-level container for everything you build — agents, integrations, secrets, members, billing. Fully isolated from other organizations. Sometimes called tenant in API contexts (the URL path uses {tenant}).

Parameter

A typed input to an agent — STRING, INTEGER, FLOAT, BOOLEAN, OPTION, JSON, STRING_ARRAY, DICTIONARY, ASSET, or TOOL. Defined under Parameters in the agent's form.

Placeholder

Template syntax for referencing data in prompts and tool configs: {{param.X}} (a parameter), {{sender.X}} and {{venue.X}} (the person and conversation a message came from), {{secret.Z}} (a secret), and <file:…> (an attached asset, inserted from the asset picker). In a workflow, {{step.Y}} references an earlier step's output.

Predefined Tool

A custom-coded integration tool (as opposed to declarative OpenAPI tools). Used when an integration needs complex C# logic.

Revision

An immutable snapshot of an agent's definition, created each time a version is saved. See Versioning.

Role

A permission level. Organization roles are Owner, Admin, Developer, Operator, or Viewer; Application collaborator roles are Viewer, Editor, or Admin. Reaching an agent requires both levels, resolved through the Application that owns it. See Team & Roles.

Sandbox Mode

A way to run an agent's integration calls without touching production. Each connection's Sandbox routing sends its calls to the vendor's own sandbox, the Fruxon simulator, or a read-through hybrid (real reads, simulated writes). A connection with no sandbox saved is still contained in a sandbox run: its writes reach production only if you choose Live (production) for it on that test. See Sandbox Mode.

Secret

An encrypted value stored in the organization and referenced via {{secret.X}}. Used for API keys, OAuth tokens, integration credentials. See Secrets.

Session

The state of a multi-turn conversation, including conversational context and any session-scoped variables. Auto-summarized as it grows. See Sessions.

Skill

A reusable bundle of instructions, tools, and knowledge that adds expertise to an agent. See Tools & Skills.

Studio

The editor where every agent is configured and tested, as one form with a test chat beside it. See Agent Studio.

Sub-agent

An agent run as part of a larger job: as an Agent step in a workflow, or as a parallel sub-task of itself. See Combining Agents.

System Prompt

The instructional message sent to the LLM that establishes the assistant's behavior, tone, and constraints.

Tag

A label attached to runs or conversations for grouping and filtering.

Tool

Anything an agent can do beyond generating text — calling APIs, querying DBs, running code, reading/writing memory. See Tools & Skills.

Trace

The full record of one agent run — inputs, prompts, tool calls, outputs, tokens, latency, errors. See Observability.

User Prompt

The user-side message in an LLM call, typically built from placeholders against the run's inputs. Labeled Input Prompt in the agent's form.

Vectorization

The process of embedding text chunks for semantic retrieval. Applied to assets when added to a knowledge base.

Workflow

A graph built on a canvas in the workflow builder: Agent steps that each run an existing agent, tool steps, and the order between them, from an entry point to an exit point. Use one when a job needs several agents and tools in a set order. An agent has no graph. See Combining Agents.