This project provides a cloud-native orchestrator for AI agents, built on Kubernetes. It addresses the need for managing long-lived, outer-loop agents that require asynchronous execution of LLM inferences and tool calls, including human feedback loops. The primary benefit is simplified, durable, and reliable agent scheduling and execution for developers building sophisticated AI applications.

How It Works

ACP operates as a Kubernetes operator, defining AI agent components—LLMs, Agents, Tools, and Tasks—as Custom Resource Definitions (CRDs). It implements infrastructure-level asynchronous execution, akin to async/await, enabling agents to pause execution during long-running tool calls or delegations and resume reliably from checkpoints. This approach ensures durability and simplifies state management by leveraging the rolling context window accumulated through interactions.

Quick Start & Requirements

• Primary Install/Run: Deploy the operator using kubectl apply -f https://raw.githubusercontent.com/humanlayer/agentcontrolplane/refs/heads/main/acp/config/release/latest.yaml.
• Prerequisites: kubectl, kind (for local Kubernetes), Docker, and API keys for LLM providers (e.g., OpenAI).
• Links: Discord, Documentation, Examples.

Highlighted Details

• Kubernetes-Native Orchestration: Manages AI agent lifecycles and resources declaratively via Kubernetes CRDs.
• Durable Agent Execution: Provides infrastructure-level checkpointing and resumption for asynchronous tasks and tool calls.
• Human Interaction Workflows: Supports both human approval for tool execution and incorporating humans as tools for non-deterministic input.
• Extensibility: Agents, tools, and tasks are defined in YAML, facilitating easy customization and sharing.

Maintenance & Community

The project is currently in alpha. Community engagement is encouraged via their Discord. Contributions in the form of issues, documentation, and pull requests are welcomed.

Licensing & Compatibility

Licensed under the Apache 2.0 License, which is permissive for commercial use and integration into closed-source projects.

Limitations & Caveats

As an alpha project, ACP may be unstable and subject to breaking changes. Key features such as improved MCP scheduling, delegation to sub-agents, and the human-as-a-tool functionality are still in progress. A dedicated UI and CLI are also planned for future releases.