How AGNT works, end to end

AGNT is a multi-tenant AI agent platform. Consumers talk to a personal agent over WhatsApp, Telegram, or Instagram. Businesses run a venue agent that answers questions and accepts bookings in sixteen-plus languages. The two agent types talk to each other over a signed envelope protocol called A2A, routed by an in-house gateway called ClawPulse. Every booking settles through a Commerce Ledger and a Stripe payout. The rest of this guide is the same story with details.

A user sends a WhatsApp message: 'any good sunset spots for Friday at 7?'. The AGNT webhook endpoint receives the inbound event from the WhatsApp Business Platform (via 360Dialog), authenticates the payload, and routes it to the soul loader.

The soul loader is the prompt builder. It pulls the user's structural memory keys (diet, favourite areas, typical party size, fitness goal), runs a pgvector cosine similarity search for relevant semantic facts, adds a time-of-day block, and assembles the system prompt. This all happens in under 80 milliseconds.

The LLM gateway receives the prompt and dispatches to Claude Sonnet 4.6 for Pro tier, or Haiku 4.5 for Free and Starter. Vision tasks always pin to Sonnet. The model calls a tool — in this case search_venues — and the tool executor runs the actual query against PostgreSQL with pgvector, returning the top matches.

The model composes the response, the messaging layer relays it back to WhatsApp, and the user gets a reply within a couple of seconds.

When the user says 'book me a table at La Brisa for 7:30', the consumer agent emits an AGPEnvelope with intent booking.search to La Brisa's venue agent through the ClawPulse gateway. The envelope carries the party size, the requested time, and an agent_context block for correlation.

ClawPulse validates the signature (HMAC-SHA256 with timestamp), checks the per-venue circuit breaker, and dispatches the envelope. The venue agent responds with one or more slots. The consumer agent picks one and sends booking.confirm, which creates a row in the Commerce Ledger in the pending state.

If the venue is not on the AGNT network, the consumer agent falls back to a public booking endpoint or generates a WhatsApp deep link. The goal is always a confirmed booking — the path is negotiable.

Everything persistent lives in one PostgreSQL 16 cluster. Users, venues, bookings, the Commerce Ledger, and every VenueChunk of a Knowledge Pack share the same relational database. pgvector extends PostgreSQL with a vector type and ANN search, so 1536-dimensional OpenAI embeddings sit right next to the relational row they describe.

Redis handles everything that does not need to persist — Fernet-encrypted session tokens, slowapi rate limit counters, distributed locks for APScheduler leader election, and the send queue that retries outbound messages with exponential backoff.

Memory is split into two shapes. Structural keys are a fixed list (diet, interests, last_booking, fitness_goal, preferred_booking_hour) always included in the system prompt. Semantic memory is everything else — recalled per conversation through pgvector.

Every confirmed booking writes a Commerce Ledger entry with three cents figures: gross (what the guest paid), platform net (AGNT's cut), and venue net (what the venue receives). The fee is either seven percent with a $1.50 floor for subscribed venues, or a flat $2.50 for unsubscribed venues.

Stripe handles settlement. An APScheduler job batches the ledger nightly, posts usage records for metered API billing ($0.25 per booking.confirm on the Open Network), and queues payouts. Refunds and disputes flow back through the ledger as inverse entries so the books stay auditable.

Anthropic Claude is the primary LLM. Haiku 4.5 for Free and Starter, Sonnet 4.6 for Pro and all vision work. OpenAI's text-embedding-3-small produces the vectors. Ollama runs llama3:70b-q4_K_M as a local fallback when Anthropic is degraded.

AI is not the whole product. AGNT is half AI (the conversation and vision layers) and half boring infrastructure (PostgreSQL, Redis, Stripe, APScheduler, HMAC-signed A2A). The AI gets the user interested; the infrastructure gets them seated.