Icepick is a TypeScript library for building fault-tolerant, scalable AI agents, targeting developers who need to integrate AI capabilities into existing codebases. It simplifies durable execution, queueing, and scheduling, allowing users to focus on core business logic rather than infrastructure complexities.

How It Works

Icepick leverages Hatchet, a durable task queue, to provide automatic checkpoints for agent execution. This ensures agents can recover from failures or wait for external events without consuming resources. Agents and tools are defined as simple functions, promoting code-first development and easy integration. The library handles distributed execution, rescheduling on machine failure, and provides configurable options for retries, rate limiting, and concurrency.

Quick Start & Requirements

• Install CLI: pnpm i -g @hatchet-dev/icepick-cli
• Create project: icepick create first-agent
• Documentation: https://icepick.hatchet.run

Highlighted Details

• Code-first agent definitions for seamless integration.
• Distributed execution across a fleet with graceful scheduling and rescheduling.
• Configurable options for retries, rate limiting, and concurrency control.
• Supports deployment on various container platforms like Kubernetes, AWS ECS, and GCP Cloud Run.

Maintenance & Community

• Community support via Discord: https://hatchet.run/discord
• Follow on Twitter: @hatchet-dev

Licensing & Compatibility

• License: MIT. Compatible with commercial and closed-source applications.

Limitations & Caveats

Icepick is not a comprehensive AI framework; it focuses on the infrastructure layer and expects users to manage LLM calls, prompts, and agent memory implementation.