Introduction

​

What is Bedrock?

Bedrock handles the hard parts of building AI agents:

• Persistent State — Agents use Bedrock-owned adapter state for contacts, messages, documents, projects, and notifications
• Tool Execution — Connect agents to SMS, email, calendars, or your own APIs via webhooks
• Autonomous Runtime — Agents run independently, waking and sleeping on schedules
• Full Observability — Every LLM call, tool invocation, and decision is traced

See a demo video here.

​

Core Architecture

• Organization — your top-level tenant. Owns every template, agent, adapter, and API key. Holds the org-wide OpenAI / Anthropic keys used when Bedrock talks to a provider.
  • AgentTemplate — a named blueprint: system prompt, default model, adapter set + per-adapter config, harness git ref, tool call secret.
    • Agent 1, Agent 2, … — created from the template. Each agent snapshots the template’s configuration + adapter configs at creation time. Template edits don’t retroactively change live agents.

Adapters and tools live on the organization and are attached to templates via AdapterConfig rows; each agent gets its own immutable AgentAdapterConfig snapshot.

​

Key Concepts

• Organization — Tenant account. Holds users, billing, LLM provider keys, and API keys.
• AgentTemplate — Blueprint for agents: prompt, model, adapter set, tool-call secret. Think of it like a Docker image.
• Agent — A running instance created from a template (or ad-hoc) with its own adapter state. Carries a purpose (production | eval | dev) and free-form tags.
• Adapter — An integration module (SMS, Email, CRM, etc.) that provides tools.
• Tool — A callable function an agent can invoke, either built-in or a webhook to your API.

​

Built-in Adapters

Bedrock includes these default integrations:

​

Custom Integrations

Build your own adapters to connect agents to any system:

1. Create an adapter — a named container for related tools
2. Add webhook tools — each tool POSTs to your API when called
3. Verify requests — check the X-Agent-Secret header to authenticate
4. Return data — your JSON response becomes the tool result the agent sees

See Custom Adapters for a full walkthrough.

​

How Agents Run

Every run spawns an isolated bedrock-harness subprocess (pinned to the template’s harness_git_ref, or latest main if unset):

1. Wake agent (API call or scheduled)
2. Load memory context
3. Build prompt with tools
4. Call LLM (OpenAI / Anthropic)
5. Execute tool calls
6. Log to memory
7. Repeat until agent sleeps

Agents are autonomous — they decide when to sleep and what actions to take based on their tools and memory.

​

Use Your Own LLM Keys

Bedrock uses your LLM provider accounts:

• You control costs directly with OpenAI / Anthropic
• Use any model available on your account
• Prompt caching with Claude reduces costs significantly

Keys live on the Organization (openai_api_key / anthropic_api_key). If unset, Bedrock falls back to OPENAI_API_KEY / ANTHROPIC_API_KEY environment variables.

​

Quick Links