Glossary
42 terms across 7 categories. The language of AI agent commerce — defined by the platform that ships it.
Protocols and Standards
Communication and discovery standards used across the AGNT network.
A2A Protocol
Agent-to-Agent communication standard that lets AI agents negotiate, book, and transact directly. Uses AGPEnvelope JSON messages with protocol version, sender, intent, and payload. Supported intents: network.ping, booking.search, venue.list, booking.confirm.
AGPEnvelope
Agent Gateway Protocol message shape. The canonical JSON format for all A2A communication. Fields: protocol, from_agent, to, intent, payload, agent_context, ttl_ms, timestamp.
MCP (Model Context Protocol)
Protocol for exposing tools to LLMs. AGNT's MCP server at /mcp/sse exposes venue search, venue listing, and booking confirmation as standard tools callable by Claude, GPT, or any MCP-compatible model.
Agent Card Standard
Machine-readable JSON identity published at /.well-known/agent-card.json. Declares an agent's name, capabilities, supported protocols, and endpoints — enabling automatic discovery by other agents and services.
x402 Protocol
Crypto payment protocol for agents enabling trustless instant micropayments between AI agents. Uses USDC on Base L2 for settlement without a human-operated payment gateway.
llms.txt
AI discovery standard that lets venues declare what they offer to language models. Replaces keyword SEO with structured intent data, making venues findable by AI agents.
Network and Infrastructure
Core infrastructure powering the AGNT agent ecosystem.
AGNT Open Network
Public A2A marketplace where external agents query venues, confirm bookings, and subscribe to webhooks. Accessible via REST API, A2A protocol, and MCP server with metered pricing.
ClawPulse
AGNT's agent intelligence network and A2A gateway. Routes booking envelopes between consumer agents and venue agents with HMAC-signed delivery, global and per-venue circuit breakers, and response listening.
Circuit Breaker
Failure detection pattern protecting the A2A gateway. Opens after 5 failures in a 10-minute window, blocking requests for 5 minutes. Half-open state allows one probe request to test recovery. Applied globally and per-venue.
Venue Graph
The connected database of 134+ venues across Bali — restaurants, bars, gyms, and spas — with real-time availability, menus, pgvector embeddings for semantic search, and booking capabilities.
Webhook
HTTP POST notification sent to developer endpoints when events occur (e.g., booking.confirmed). Each delivery is HMAC-signed via X-Agnt-Signature header. Supports retry, replay, and secret rotation from the developer dashboard.
Agents and Intelligence
AI agent types, identity, and reasoning concepts.
Personal Agent
AI assistant that lives in a user's WhatsApp, Telegram, or Instagram. Handles venue search, bookings, calorie scanning, product price comparison, transport, and reminders. Gets smarter over time through memory and preference learning.
Venue Agent
AI agent representing a business (restaurant, bar, gym, spa). Answers customer questions, handles booking requests, and communicates in 16+ languages. Built from the venue's knowledge pack and Venue Soul.
Venue Soul
The AI model of a venue's identity — voice, vibe, rules, and recommendations — that underpins every response the venue agent gives. Built from knowledge extraction of menus, policies, and FAQs.
Soul Loader
System prompt builder that constructs the LLM context from user memory (structural keys + semantic recall), time-of-day context, and recent activity. Cached in Redis for 1 hour, invalidated on memory writes.
AI Concierge
An AI-powered booking and discovery assistant that handles guest requests, recommendations, and reservations in the guest's own language via WhatsApp or Telegram. Another name for a personal agent with hospitality focus.
External Agent
Non-AGNT AI system integrating via A2A protocol or MCP. Has its own agent_id, API key, and tier. Interactions tracked separately with source: external_agent.
Commerce and Booking
Booking lifecycle, payments, and revenue mechanics.
Booking Confirmation
A2A intent (booking.confirm) that creates a reservation on the AGNT network. Available only on pro tier at $0.25 per confirmation. Status follows a state machine: pending, confirmed, completed, cancelled, no_show.
Referral Fee
Platform commission from bookings. Default 7% of booking value or $1.50 floor for subscribers. Flat $2.50/booking for unsubscribed venues. Tracked in the Commerce Ledger.
Metered Billing
Per-use pricing for API consumers. Each booking.confirm call costs $0.25, tracked in ApiKeyUsageEvent and batched nightly to Stripe usage records. No upfront fees or monthly minimums.
Commerce Ledger
Financial transaction log tracking every booking fee. Records gross, platform net, and venue net amounts in USD with status: pending, charging, settled, or failed. Linked to Stripe payment references.
Deposit
Optional pre-payment for venue bookings stored in cents. Linked to a Stripe PaymentIntent. Reduces no-show risk for high-demand venues.
Memory and Knowledge
How AGNT agents learn, remember, and reason.
User Memory
Fact store for each user containing key-value pairs with confidence scores (0.0-1.0) and pgvector embeddings. Structural keys (diet, interests, birthday) are always included in prompts. Semantic keys are recalled via cosine similarity.
Semantic Recall
Embedding-based memory retrieval. User message is embedded, then searched against user_memory via pgvector cosine distance. Returns top 10 non-structural facts. Skipped on trivial messages (greetings, emoji) to save cost.
Knowledge Pack
Venue documentation bundle — menus, policies, hours, FAQs — ingested and chunked into VenueChunk records with pgvector embeddings for RAG retrieval by the venue agent.
Knowledge Extraction
Automated ingestion of menus, policies, hours, and FAQs from a venue's existing documents. Turns unstructured text into structured agent memory chunks with vector embeddings.
Structural Keys
Memory keys always included in the agent's system prompt: diet, fav_areas, interests, last_booking, typical_party_size, preferred_booking_hour, fitness_goal, daily_calories, dietary_restrictions, and more. Never require embedding search.
Developer Concepts
API, authentication, and integration concepts for builders.
API Key
Credential for accessing the AGNT Open Network. Format: agnt_live_[32-char-hash]. Stored as SHA-256 hash with visible key_prefix. Has tier (free/pro), scope, daily quota, and usage counters. Raw key shown once at creation.
Rate Limiting
Request throttling per API key. Free tier: 100 requests/day. Pro tier: 10,000 requests/day. Limits reset at midnight UTC. Exceeding returns HTTP 429 Too Many Requests.
HMAC Signing
Message authentication for A2A and webhooks. Payload + timestamp signed with the A2A signing key or webhook secret using HMAC-SHA256. Prevents message tampering. Signatures valid for 5 minutes.
Intent
Named action in the A2A protocol. Each AGPEnvelope carries an intent that routes to a specific handler. Supported: network.ping, booking.search, venue.list, booking.confirm.
Token Budget
Per-tier daily LLM token limit. Free: 50K tokens/day. Starter: 200K. Pro: 1M. Tracked in Redis with 24-hour expiry. Prevents runaway costs from individual users.
Ecosystem Tools and Surfaces
External AI tools and protocols AGNT composes with — the connective layer for the modern AI tool stack.
MCP Server
A process that exposes tools to LLMs via the Model Context Protocol. Declares a set of tools with input/output schemas; any MCP-compatible client (Claude Code, Codex, custom clients) can call them without writing provider-specific adapters. AGNT runs an MCP server at /mcp/sse exposing venue search, memory, and knowledge graph.
MCP Host
An environment that consumes external MCP servers and surfaces their tools to a model. Claude Code is an MCP host. AGNT is also an MCP host — it can register external MCP servers and expose their tools to AGNT agents at runtime.
Claude Code Skill
A reusable prompt + instruction bundle for Claude Code. Declares a name, description, triggering heuristics, and instructions. Claude Code routes into the skill when its description matches the user's intent and the instructions become part of the active prompt. Forwards cleanly into AGNT's soul loader.
Plugin
An installable bundle containing skills, subagents, commands, hooks, and/or MCP servers. Plugins are portable across sessions, teams, and model swaps — a well-shaped plugin is agent-runtime-agnostic. AGNT consumes plugin-shaped bundles through Paperclip.
Hermes Router
An open-source, model-agnostic agent router with strong function-calling support. Classifies intents and routes between models and tools. AGNT can forward Hermes routing decisions into the heartbeat stream for supervisor-level observability.
OpenClaw Protocol
Open protocol for agent-to-agent commerce. Defines envelope shapes, intent vocabulary, and trust primitives. AGNT's ClawPulse gateway is a reference implementation of the OpenClaw envelope protocol, extended with commerce-grade primitives (venue availability negotiation, x402 settlement, 15-second round-trip cap).
Fleet Adapter
A narrow, swappable interface that wraps a model backend in AGNT's fleet layer. Four methods: complete, stream, supports_tools, health_check. Adapters exist for claude_local, codex_local, gemma_local, and gemini_local. Swaps are hot — no agent restart, cwd preserved across the swap.
Soul Loader Module
A declarative capability unit registered with AGNT's soul loader. Cached per agent for 1 hour, invalidated on update. Claude Code skills forward into AGNT as soul-loader modules — same markdown, two surfaces.
Tool Description
The text string an MCP tool declares as its description. The single biggest lever for routing accuracy — LLMs decide whether to call a tool based largely on the description. First 10 words must include the action verb and primary object. Tighten every description before shipping.
Heartbeat Event
A structured JSON event emitted by every AGNT agent describing what it did, what tool it called, and what it decided. Replayable in the supervisor UI. Hermes routing decisions forward into heartbeat as first-class events, making routing observable alongside execution.
FAQ
Common questions
Quick answers about AGNT terminology.
A2A (Agent-to-Agent) is AGNT's envelope-based protocol for direct inter-agent communication. Agents exchange AGPEnvelope JSON messages containing protocol version, sender, intent, and payload. It enables AI agents to search, book, and transact without human-operated APIs.
Start building on AGNT.
Now you know the language. See it in action.