Introduction

A3T is an agent infrastructure platform built on FlowR, a CNCF Serverless Workflow runtime written in Rust.

What it does

A3T lets you build, deploy, and operate AI agents as workflows. Instead of writing application code, you define agent behavior as YAML workflows that the FlowR runtime executes.

steps:
  - name: classify
    type: call:llm
    resource: claude-sonnet
    config:
      prompt: "Classify this ticket: ${ .input.text }"
    then: route

  - name: route
    type: call:http
    config:
      method: POST
      endpoint:
        uri: https://api.internal/tickets
      body: "${ .classification }"

Core components

Component	Purpose
FlowR	Workflow runtime — executes CNCF Serverless Workflows
ZeroClaw	Agent framework — LLM reasoning + tool use + memory
ClawDB	Agent memory — 5 memory types via AQL
A3T Console	Management UI — templates, agents, canvas, resources

Key concepts

• Workflows — YAML definitions with steps, each step is a task type
• Task Types — 26 built-in types: call:http, call:llm, call:agent, call:lancedb, switch, for, etc.
• Resources — Connections to external systems: LLMs, databases, MCP servers
• Agents — LLM-powered entities with memory and tool access, running as workflows
• Templates — Versioned, reusable agent blueprints
• Canvas — Multi-agent orchestration specs
• Triggers — Webhooks, cron schedules, and events that start workflows