agentcontrolplane

Agent Control Plane (ACP) is a cloud-native orchestrator for AI Agents built on Kubernetes. It enables long-lived outer-loop agents that can perform asynchronous execution of both LLM inference and long-running tool calls, with strong durability and reliability guarantees. ACP models its design around the 12-factor principles for agents that contact humans or delegate work to other agents, making it suitable for deploying AI workflows that require orchestration across tools, humans, and sub-agents. To use ACP, you typically deploy the ACP operator into a Kubernetes cluster and then create resources such as LLMs, Agents, and Tasks to drive your AI workflows. The system is intended to be alpha, so expect ongoing improvements and evolving APIs as the project matures.

In practice, you define an LLM resource pointing to a provider (for example, OpenAI) with the appropriate API keys, then create an Agent resource that references the LLM and provides a system prompt. Tasks can then be submitted to interact with the agent, prompting the agent to perform steps that may include calling tools, delegating to sub-agents, or requesting human input. Tools in ACP are represented as MCP Servers, Humans, or other Agents, enabling rich, multi-actor workflows. The documentation also covers how to add tools via MCP, how to work with various language models, and how to incorporate human approval into decision-making processes.

Prerequisites:

• A Kubernetes cluster (local or cloud) and kubectl configured to talk to it
• kubectl installed on your machine
• Docker (for local image builds) or access to a container registry

Step 1: Install the ACP operator and CRDs

• Deploy the ACP operator and CRDs from the project repository:

  kubectl apply -f https://raw.githubusercontent.com/humanlayer/agentcontrolplane/refs/heads/main/acp/config/release/latest.yaml
  

Step 2: (Optional) Install a specific version of ACP components

• If you want to install a specific version of the operator and CRDs:

  kubectl apply -f https://raw.githubusercontent.com/humanlayer/agentcontrolplane/refs/heads/main/acp/config/release/v0.1.0.yaml
  

Step 3: Verify installation

• Check your CustomResourceDefinitions and pods in the acp namespace (or the namespace you install into).

  kubectl get crds | grep acp.humanlayer.dev
  kubectl get pods -n default
  

Step 4: Create and apply your ACP resources

• Define LLM, Agent, and Task resources as described in the README under Getting Started, then apply them with kubectl apply -f -.

Notes:

• The ACP workflow relies on Kubernetes resources rather than a standalone server process. The operator watches for LLM/Agent/Task CRs and runs the orchestration accordingly.
• Ensure your OpenAI API key (or alternative provider credentials) is available as a Kubernetes secret and referenced properly in your LLM resource.

Tips and common issues:

• If you don’t see Ready status, check the operator logs and the events for the Agent or LLM resources to identify validation errors or missing credentials.
• When running locally, ensure your Kubernetes cluster has sufficient resources to manage the ACP components and any long-running tasks.
• For different language models, verify provider-specific parameters in the LLM spec (e.g., model name, parameters).
• If tool calls fail, confirm the MCP Server or tool resource definitions (Tools) are correctly created and accessible by the ACP components.
• Review the ACP documentation and examples folder for sample LLM/Agent/Task configurations to accelerate setup.