Glossary

Definitions of key platform terms - from acceptance regions and actions to world model and workspaces.

Key terms used throughout the Amigo platform documentation. Definitions are kept short and practical. Terms are listed in alphabetical order.

---

Acceptance Region: The set of outcomes that count as successful for a given use case. Defined across multiple dimensions (accuracy, safety, empathy, latency, cost). An outcome must satisfy all dimensions simultaneously to fall within the acceptance region.

Action (API: tool): A program or integration that an agent can execute during a conversation. Actions connect the agent to external systems such as scheduling APIs, EHR lookups, or notification services. Called "tool" in the API.

Agent: A conversational AI system configured to handle a specific set of tasks. Agents have a defined persona, context graph, dynamic behaviors, and evaluation criteria.

Agent Forge: The CLI tool for managing agent configurations. Used to create, update, version, and promote services and version sets across environments. See the Agent Forge reference page.

Audio Embedding: A dense vector representation of an audio segment that captures paralinguistic features (tone, urgency, hesitation) without transcription. Enables semantic search over how something was said, not just what was said.

Audio Intelligence: A real-time verification layer that cross-checks STT output using a separate LLM during live calls. Catches misrecognized medical terms and proper nouns before they reach the agent's reasoning pipeline. Distinct from post-call re-transcription, which runs after the call ends.

Auto-Enrichment: Background process that generates vector embeddings for world model events after they are written. Runs asynchronously with zero impact on write latency. Enables semantic search over clinical data.

Backfill: The process of replaying historical conversation data through an updated configuration to regenerate metrics and verify that changes produce the expected improvements.

Barge-In: The ability for a caller to interrupt the agent while it is speaking. The voice pipeline detects incoming speech during TTS playback and stops the agent's audio so the caller can be heard immediately.

Burst-to-Experience Mapping: The translation layer between vocal burst classifications (sighs, laughs, groans) and TTS emotion parameters. Maps 25 vocal burst types to voice delivery adjustments so the agent's tone responds to non-speech sounds.

Call-Phase Adaptation: Automatic adjustment of agent behavior based on call duration and emotional trajectory. Prevents calls from dragging on when the caller is clearly unhappy.

Clinical Tools: The 13 built-in tools available to the voice agent during live calls. Cover patient lookup/create/update, appointment scheduling/cancellation/confirmation, insurance creation, outbound call scheduling, and semantic search. All write operations fire outbound sync events through the connector runner.

Confidence Score: A numeric value used in two contexts. (1) Data confidence ranks information by source reliability: 1.0 (authoritative system integration), 0.7 (browser scrape), 0.5 (voice/conversation extraction), 0.3 (agent inference), 0.0 (rejected/contradicted). Higher-confidence sources overwrite lower ones in the world model. (2) Agent confidence measures how certain the agent is about its current response or decision. When agent confidence drops below a configured threshold, the agent escalates to a human operator.

Conference-First Architecture: The call transfer model where the agent, caller, and operator are bridged into a three-way conference before the agent drops off. This ensures the operator hears the conversation context and the caller experiences a smooth handoff rather than a cold transfer.

Connector Runner: The bidirectional integration layer that loads external data into the platform and syncs verified data back to external systems.

Context Graph (API: service_hierarchical_state_machine): The structured map of a problem space that guides agent behavior. Context graphs define states, transitions, decision points, and safety boundaries for a specific workflow. Called "service hierarchical state machine" in the API.

Conversation Starter: An LLM-generated opening prompt personalized to the agent's persona, the user's model, and the current service context. Used to help users initiate conversations in the Classic API.

Data-MCP: A Model Context Protocol server that exposes workspace data through SQL query tools. Compatible with MCP clients such as Claude Desktop and Claude Code. Provides seven tools for querying, profiling, and exploring workspace tables.

Drift: A gradual change in agent performance over time. Can be caused by changes in the input distribution (new types of conversations), shifts in agent behavior, or evolving requirements. See Drift Detection.

Dynamic Behavior (API: dynamic_behavior_set): A behavior that activates at runtime based on conversational context, user signals, or external events. Dynamic behaviors let the agent adapt its approach in real time. Called "dynamic behavior set" in the API.

Entity Resolution: The process of matching records across multiple data sources to a single patient or user identity. Used by the connector runner to unify data from different systems.

Escalation: The process of handing a conversation from the agent to a human operator. Escalation is triggered by low confidence, safety boundary conditions, or explicit patient requests.

File Drop Connector: A connector type that ingests data from files (CSV, NDJSON, FHIR Bundle, JSON) deposited in an S3 bucket. Used for batch data imports from partners or external systems.

Filler Speech: Short spoken phrases (such as "Let me check on that" or "One moment") that the agent produces while waiting for a backend operation to complete. Filler speech prevents dead air during tool calls and keeps the caller engaged.

Interaction Insight: A detailed view of the agent's reasoning for a specific interaction, including which memories were active, what reflections were generated, which state transitions occurred, and which tools were considered. Used for auditing and debugging agent behavior.

Keyterm Boosting: A speech-to-text configuration that increases recognition accuracy for domain-specific vocabulary. Medical terms, drug names, provider names, and other specialized words are added to a boost list so the STT engine favors them over phonetically similar common words.

LLM Translation: The process of using an LLM to translate world model entity state into EHR-specific formats during outbound write-back. Falls back to deterministic mapping when the LLM is unavailable.

Metric: A configured evaluation criterion that measures a specific dimension of agent performance. Metrics can be scored automatically after each conversation or during simulation runs. See Metrics.

Monitor Concept: A tracked signal or condition that the platform watches for across conversations. Monitors can trigger alerts, escalations, or dynamic behavior changes when specific patterns are detected.

Operator: A human staff member who receives escalated conversations from the agent. Operators get a handoff summary with conversation context and patient data so they can continue without starting over.

Outbound Task: An entity type in the world model that represents a scheduled outbound call. Created by scheduling rules, follow-up workflows, or manual triggers. The connector runner's outbound dispatch loop reads these entities and initiates calls when they are due.

Patient Agent: An LLM-powered agent within the connector runner that autonomously creates patient records in EHR systems when a new patient is identified during a voice call. Runs on HIPAA-compliant infrastructure.

Persona (Simulation): A synthetic user profile used in simulations. Defines the characteristics, behaviors, and communication style of a test user. See Simulations.

Proactive Emotional Intelligence: The voice agent's ability to detect sensitive topics from context graph content and adjust delivery before the caller reacts. Prevents blunt delivery of difficult information.

Projection Function: The function that computes an entity's current state from its events. Different entity types have different projection functions (patient, outbound task, generic). Runs inside the same database transaction as the event write.

Quality Score: A 1-5 rating produced by post-call analysis across five dimensions: task completion, information accuracy, conversation flow, error recovery, and caller experience.

Regulation Template: A configurable compliance framework that encodes regulatory requirements (HIPAA, state-specific rules, organizational policies) as constraints the agent must follow.

Review Queue: The human review interface where operators examine clinical events flagged by the automated pipeline. Supports approve, reject, and correct actions with confidence elevation.

Risk Score: A composite per-turn assessment of call health in the voice agent. Combines emotion signals, loop detection, and call duration into a score that maps to four levels: normal, monitor, alert, and escalate.

Safety Triage: Per-turn regulatory safety evaluation that runs against pre-built templates (Joint Commission NPSG 15, VAWA, FDA MedWatch). Returns concern levels 0-3 independent of monitor concept detection.

Scenario (Simulation): A defined situation used in simulations. Describes the context, events, and goals for a test interaction. See Simulations.

Silence Monitor: The voice agent component that detects and manages caller inactivity. Uses exponential backoff check-ins (10s, 20s, 40s) before ending the call with an operator transfer offer.

Skill: A discrete capability that a workspace exposes through the Platform API. Skills define what the agent can do in a specific context, such as scheduling appointments or conducting symptom assessments.

Test Run: The execution of a test set that produces scored results for each unit test. See Simulations.

Test Set: A group of related unit tests that are run together. Test sets are often organized by capability area or risk level.

Tone Momentum: A caching mechanism in the TTS engine that preserves the previous turn's emotional parameters. Prevents jarring vocal shifts when the emotion detection signal is temporarily weak or unavailable.

Transcript Extraction: The process of pulling structured patient data (phone, DOB, email, insurance, address) from conversation transcripts during or after a call. Extractions are written to the world model with voice-level confidence.

Triage Hint: A linguistic or behavioral pattern included in a safety template that tells the LLM what to watch for beyond direct statements. Examples: farewell language for suicide risk, unexplained injuries for domestic violence.

Unification Engine: The transformation layer that converts raw records from any connector type into world model events using configurable mapping rules with dot-path field extraction.

Unit Test: A combination of a persona, a scenario, and success criteria that tests a specific agent behavior.

Version Set: A named collection of component versions (agent, context graph, dynamic behaviors) that are deployed together. Version sets are promoted through environments (staging to production) as a unit.

Webhook Connector: A connector type that receives push-based data from external systems via HTTP webhooks, rather than polling. Events are deduplicated by content hash.

World Model: The platform's unified data layer that assembles information from multiple sources (EHR, conversations, manual entry, connected devices) into a single view accessible to the agent during conversations.

Workspace: A container in the Platform API that groups related skills, agents, and configurations for a specific deployment context.

Write Scope: A permission boundary that limits what entities and event types the voice agent can modify during a call. System services bypass this restriction. Prevents voice-extracted data from overwriting authoritative records.