Introduction

​

The Problem

1. AI agents are useless without reliable tools.

There’s already a lot of code in your internal systems, but it’s not easy to make tools out of them.

• If your code is distributed across multiple services, you have to build APIs for them.
• To make your tools reliable and efficient, you need to have retries, fail-over (to a replica), load balancing, caching, etc.

2. AI agents perform best when they have a single, well-defined task.

This means, you need a reliable way to orchestrate multiple agents to achieve a single goal.

• AI agents and their worflows can run for a long time, and need to be durable.
• Without workflow as code, you have to build a UI to define your workflows. This is error-prone, and becomes an unmaintainable mess for complex workflows.
• You need a way to version your workflows, so you can update them without affecting existing executions.

3. Your production agents and tools live in a distributed environment.

Your tools and agents might be written in different languages, run in different cloud environments. In order to make them work together, you need a way to reliably invoke and orchestrate them.

• You need a way to trigger your workflows, and invoke your tools without exposing your internal systems to the outside world.
• AI agents and tools need durable execution, so they can survive long-running processes, interruptions, and failures.

​

Our Solution

Inferable tries to solve these problems by providing a vertical stack that brings tools, agents, and workflows (durable execution) together.

1. Language SDKs to wrap your existing code into tools.

Multiple language SDKs allow you to convert your existing code into tools, or build new tools from scratch.

• Wrap your existing code in a tool.
• Developer primitives to put human in the loop.
• Supports running multiple copies of your tools in parallel, and to load balance between across replicas.
• Supports caching, retries, and fail-over.

2. Workflow as code and durable execution engine.

Workflow as code allows you to define agents interactions as code, and orchestrate them as code on top of a durable execution engine.

• Define your workflows as code, version them, and run them from anywhere.
• Agent interactions are defined using existing programming primitives, like if, else, for, while, etc. Nothing new to learn.
• Agents are durable, and can survive long-running processes, interruptions, and failures.

3. Support for distributed environments without infrastructure complexity.

The control plane is a single, reliable source of truth for your workflows, agents and tools.

• Triggering workflows, tools and agents from anywhere, using code, or a HTTP request.
• Long-polling means you don’t have to expose your internal systems, tools and agents to the outside world, or build an API for them.
• Monitoring and observability of your workflows, agents and tools.