Glossary

Definitions of terms used throughout Cruvero documentation.

Agent — An LLM-powered entity that executes a task through a series of observe-decide-act steps within a Temporal workflow.

Agent State — The mutable runtime state of an agent including conversation history, tool results, and metadata, stored in Temporal workflow state.

Anomaly — A recorded instance of unexpected tool behavior detected by the immune system, characterized by a signature hash and hit count.

Approval — A human-in-the-loop checkpoint where an agent pauses execution and waits for explicit human authorization before proceeding.

Blackboard — A shared key-value store enabling inter-agent communication within supervisor workflows.

Causal Trace — A per-step record linking each agent decision to its inputs, observations, and outcomes for debugging and forensic analysis.

Compose Tool — A meta-tool that chains multiple tool calls into a single composite operation with defined steps and argument mappings.

Context Budget — A configurable token limit that controls how much memory and context an agent includes in each LLM decision prompt.

Decision — The LLM output at each agent step, selecting an action (tool call or halt) with arguments.

Decision Log — The ordered sequence of all decisions made during an agent run, used for replay and comparison.

Embedding — A dense vector representation of text, used for semantic similarity search in memory and tool discovery.

Failover Chain — An ordered list of LLM providers that the system cycles through when the primary provider fails or exceeds latency thresholds.

Golden Log — A recorded agent execution trace used as a test fixture for deterministic replay comparison.

Graph Workflow — A DAG-based workflow where nodes represent agent tasks and edges represent data dependencies.

Immune System — The automated monitoring subsystem that detects anomalous agent behavior and can quarantine tools or halt runs.

MCP — Model Context Protocol, a standard for connecting LLM applications to external tool servers.

Memory — The episodic and semantic memory subsystem that stores and retrieves context from past agent interactions.

Metacognitive Monitor — The Phase 10A subsystem that tracks agent reasoning quality and detects cognitive patterns like loops or confusion.

Observation — The result returned to an agent after a tool execution or system event.

Playground — A sandboxed environment for testing agent configurations with simulated tools and deterministic scenarios.

Provenance — A tamper-evident record of every action, decision, and data dependency in an agent run, stored as a DAG.

Quarantine — The state of a tool that has been disabled by the immune system due to detected quality degradation or anomalous behavior.

Quota — Per-tenant usage limits controlling requests per minute, requests per hour, tokens per day, and cost per period.

Registry — See Tool Registry.

Salience — A scoring mechanism that determines the relevance of memory entries to the current agent context based on weighted factors (relevance, recency, confidence, usage).

Scaffold — A generated project template containing pre-configured files for a new agent project.

Supervisor — A multi-agent orchestrator that delegates tasks to child agents and aggregates their results using patterns like broadcast, debate, or saga.

System Prompt — The initial instruction set given to an agent that defines its role, constraints, and available tools.

Temporal — The durable execution engine (Temporal.io) that provides workflow orchestration, retry logic, and state persistence for agent runs.

Tenant — An isolated organizational unit with its own agents, tools, memory, and configuration, enforced via Postgres RLS and Temporal namespace separation.

Tool — An executable capability (function, API call, or script) that an agent can invoke during a workflow step.

Tool Registry — The catalog of available tools with metadata, versioning, content hashing, and semantic search capabilities.

Vaccination — The process of marking an immune system anomaly as resolved by providing a fix procedure, preventing the anomaly from triggering further quarantines.

Vector Store — The backend (Qdrant or pgvector) that stores and queries embedding vectors for semantic similarity search.