What orchestration types does AgentBackend support?

AgentBackend supports 5 orchestration types: Single (standalone Q&A), Chain (sequential pipeline), Supervisor (manager delegates to workers), Society of Mind (multi-agent collaboration), and Workflow (directed graph with branching).

How does Agent Data Store work?

Agent Data Store gives each user an isolated PostgreSQL schema. Design schemas using the built-in Database Agent, then add tenant_query and tenant_schema_info tools to your agents so they can query structured business data like customers, orders, and inventory.

What messaging channels can I deploy agents to?

Telegram is fully supported with bot registration and webhook-based messaging. Slack integration is coming soon. WhatsApp, Discord, and web widget are on the roadmap.

Product

From prototype to production.

You've seen what AgentBackend does. Here's how it all works — orchestration, data, channels, security, and the developer experience.

ARCHITECTURE

One platform. Every layer.

AgentBackend sits between your application and the AI ecosystem — handling orchestration, data, channels, and security so you don't have to.

Your App

AgentBackend API

Model Gateway

OpenAI, Anthropic, Google, Mistral

Knowledge / RAG

Vector search, embeddings

Agent Data Store

PostgreSQL per tenant

Tools

28+ built-in integrations

Channels

WhatsApp, Telegram, Slack, REST

Guardrails

Security + observability

You manage

→Your app’s frontend
→Business logic & prompts
→User authentication
→Domain-specific agent instructions
→Your data & API keys
→When to call which agent

That's it. We handle the rest.

AgentBackend manages

✓Model routing & fallback
✓RAG pipeline & embeddings
✓Agent Data Store (PostgreSQL)
✓Tool integrations (28+)
✓Channel deployment
✓Guardrails & security
✓Observability & tracing
✓Session memory
✓Orchestration engine

ORCHESTRATION

Five patterns. One engine.

Choose the orchestration pattern that fits your use-case — from a single agent to a full directed workflow.

Single Agent

single

When to use: Simple tasks with a single focus

A standalone agent that processes requests independently. Best for Q&A, content generation, or data lookups.

Chain

chain

When to use: Sequential multi-step pipelines

A sequence of agents where each output feeds the next. Ideal for multi-step pipelines like research → summarize → format.

Supervisor

supervisor

When to use: Complex tasks requiring specialist delegation

A manager agent that delegates sub-tasks to specialist workers and synthesizes results. Great for complex, multi-domain tasks.

Society of Mind

society_of_mind

When to use: Tasks requiring diverse perspectives

Multiple agents collaborate and refine each other's outputs through structured dialogue. Best for code review, strategic analysis.

Workflow

workflow

When to use: Business processes with conditional logic

A directed graph with conditional branching, parallel execution, and merge points. Use for complex business processes with decision points.

CHANNELS

Deploy once. Reach users everywhere.

Same agent, same knowledge, same tools — every channel.

REST API

Call from any language via SDK or HTTP. Full streaming support.

Business API integration. One-click setup, webhook managed.

Bot API with automatic webhook registration. Instant deploy.

Slack

Workspace app with slash commands. Message parsing included.

Web Widget

Embed a chat widget in any website. Customizable styling.

Scheduled

Cron-based autonomous execution. Agent works while you sleep.

DATA

Three data layers. Use what you need.

Static documents for RAG, runtime context per call, and persistent storage for structured business data.

Layer	What it is	When to use	Persistence
Knowledge Base	RAG over documents	Static docs, PDFs, guides	Persistent, read-only
Context Injection	Runtime data per call	User-specific, session data	Ephemeral, per-run
Agent Data Store	PostgreSQL tables	Orders, customers, inventory	Persistent, read-write

Python SDK

response = ab.run(
    agent_id="support",
    message="What's my order status?",
    context={
        "user_id": "u_9182",
        "plan": "premium",
    },
    knowledge="product-docs",
    # Data Store accessed via tenant_query tool
)

Knowledge Base

Upload PDFs, URLs, structured data. Agent retrieves relevant chunks at query time via RAG.

Context Injection

Pass user-specific data per call. Ephemeral — never stored. Personalizes every response.

Agent Data Store

Persistent PostgreSQL tables. Agent reads and writes structured business data.

AGENT DATA STORE

Your agents get their own database.

Not just chat — your agents store, query, and act on real business data.

PostgreSQL per tenant

Every user gets an isolated schema (tenant_{uuid}). Full SQL capability.

Database Agent

Design and create tables through natural language conversation. No SQL knowledge required.

tenant_query tool

Agents can run SELECT, INSERT, UPDATE, DELETE, CREATE TABLE, ALTER TABLE on their tenant schema.

Schema-aware agents

Your data structure is automatically injected into the agent’s context. It knows your tables and columns.

Migration tracking

Every DDL change is versioned. Deterministic rollback SQL for each migration.

DDL approval

Structural changes (CREATE TABLE, ALTER TABLE) require human approval before execution.

Example flow

User“What orders shipped today?”

AgentReceives schema context (orders table with columns: id, status, date, tracking)

SQLSELECT * FROM orders WHERE status='shipped' AND date=CURRENT_DATE

Result12 rows returned from tenant_a1b2c3.orders

Agent“You have 12 orders shipped today. Here are the tracking numbers...”

SECURITY

Three layers of protection

Every message passes through input sanitization, prompt shielding, and output guardrails before reaching your users.

Context Sanitization

Incoming user messages are scanned for prompt injection attempts before reaching the LLM. Malicious instructions are stripped automatically.

System Prompt Shield

Your system prompt is never exposed to end users. Even if they ask the agent to repeat its instructions, it won’t comply.

Output Guardrails

Responses are checked for PII, off-topic content, and policy violations before being returned to the caller.

OBSERVABILITY

Step-by-step trace

Every agent run produces a full trace — from input tokenization to response streaming. Debug any issue in seconds.

#	Step	Detail	Duration
1	Input received	User message tokenized	2ms
2	RAG retrieval	3 chunks from knowledge base	48ms
3	Tool call	order_lookup → {status: shipped}	120ms
4	LLM inference	Claude 3.5 Sonnet, 312 tokens	890ms
5	Output guardrail	PII check passed	8ms
6	Response streamed	First token at 210ms	1068ms total

Start building. $3 free credit.

Zero idle cost — pay only when your agents run.

Open Console Read Docs

View SDK →