Last updated · Santosh Kumar Radha

Open-source control planes for AI agents: what to look for

What an AI agent control plane does, the open-source options in 2026, and how AgentField (Apache 2.0) runs on Kubernetes, AWS or Google Cloud.

An AI agent control plane is the shared service that registers your agents, routes calls between them, runs long work durably, applies policy and records what happened, so each agent does not rebuild that plumbing; AgentField is an open-source control plane under Apache 2.0, and the other open-source options in 2026 include kagent, Dapr Agents, Temporal and Dify, each covering a different part of the job. "Control plane" means different things to different projects, so start from what you need it to control.

At a glance

OptionWhat it controlsLicenceRuns on
AgentFieldAgents as API services: routing, discovery, async execution, memory, policy, identity, auditApache 2.0Laptop (SQLite), Docker, Kubernetes, AWS ECS, Cloud Run, Railway, Fly.io
kagentAgents defined as Kubernetes custom resources; CNCF Sandbox since May 2025Apache 2.0Kubernetes
Dapr AgentsAgents on Dapr's durable workflow engine, with retries and crash recoveryApache 2.0Wherever Dapr runs, including Kubernetes
TemporalDurable execution for any code, including agent loopsMITSelf-hosted cluster or Temporal Cloud
DifyVisual LLM app and agent builderModified Apache 2.0 (no multi-tenant service without permission; keep the logo)Docker Compose
LangSmith Agent ServerHosting for LangGraph agentsProprietary (LangGraph itself is MIT)LangChain cloud, or self-hosted on the Enterprise plan

What to look for

  1. Agents as callable services. Can another service call an agent over HTTP without importing its code?
  2. Long-running work. Queues, async execution, status and webhooks, not a request that times out at 30 seconds.
  3. Agent-to-agent routing and discovery. Calls by logical name, with health-aware lookup.
  4. Policy before execution. Rules about which agent may call which function, enforced outside the agent.
  5. Identity and audit. Who did what, provably.
  6. Licence and self-hosting. OSI-approved licence, and whether the production deployment path is open too.
  7. Model choice. Bring any model, including open-weight models you host.

How AgentField covers it

You write agents in Python, TypeScript or Go; each @app.reasoner() or @app.skill() becomes POST /api/v1/execute/<node>.<function> (see running agents as backend services). The control plane owns discovery, routing, shared memory, policy, async execution and audit. Governance is built in: tag-based access policies, W3C DIDs per agent, and signed execution credentials.

The deployment path is one binary with two storage modes:

bash
af server
# local dev, SQLite, zero extra infrastructure

export AGENTFIELD_POSTGRES_URL="postgres://user:pass@db.company.com/agentfield"
af server
# same binary, production storage backend

On Kubernetes, the docs ship a Deployment and Service for the agentfield/control-plane image with liveness and readiness probes on /api/v1/health and Prometheus metrics on /metrics. On cloud platforms:

bash
gcloud run deploy agentfield \
  --image ghcr.io/agent-field/agentfield:latest \
  --port 8080 \
  --set-env-vars "AGENTFIELD_STORAGE_MODE=postgres,AGENTFIELD_POSTGRES_URL=..."

For AWS ECS or Fargate the docs say to use the Docker image with EFS, or PostgreSQL on RDS for production. The production checklist adds: PostgreSQL, a strong AGENTFIELD_AUTHORIZATION_ADMIN_TOKEN, DID authorization on, tag approvals set to manual, and TLS at your load balancer.

Honest limits: there is no hosted AgentField; you operate it. The docs have no Azure-specific guide, so on Azure you use the same container image and Kubernetes manifests. Multi-replica control planes need PostgreSQL, because SQLite does not support concurrent writes. And kagent, Dapr and Temporal sit inside the CNCF or large communities, which gives them a bigger operator ecosystem than ours today.

When something else fits better

  • Kubernetes-first teams who want agents as CRDs. kagent defines each agent as a Kubernetes custom resource, so agents fit the cluster tooling you already use. AgentField runs on Kubernetes but does not model agents as CRDs.
  • You already run Dapr. Dapr Agents builds on Dapr and its durable workflow engine, which retries tasks and resumes after crashes.
  • Durability is the main problem. Temporal is the most proven durable-execution engine, MIT licensed, with eight SDKs.
  • Non-developers build the agents. Dify's visual builder is aimed at them; check its licence conditions if you plan to offer it as a multi-tenant service.
  • Your agents are LangGraph graphs and you want them managed. LangSmith Agent Server is the native host, on a paid plan.

For alternatives to cloud vendor platforms, see cloud agent platform alternatives. For regulated deployments, see self-hosted agents for banks and healthcare.

Questions people ask

What are the options for an open-source control plane for AI agents?

AgentField (Apache 2.0) runs agents as API services with routing, async execution, policy, identity and audit. kagent (Apache 2.0, CNCF Sandbox) runs agents as Kubernetes resources, Dapr Agents (Apache 2.0) adds agents to Dapr's durable workflows, Temporal (MIT) provides durable execution, and Dify offers a visual builder under a modified Apache 2.0 licence.

Is there an Apache 2.0 platform for running AI agents in Kubernetes?

Yes. AgentField's control plane and SDKs are Apache 2.0, and its docs include a Kubernetes Deployment with health probes on /api/v1/health; use PostgreSQL for more than one replica. kagent is also Apache 2.0 and Kubernetes-native.

Can I run AgentField on AWS, GCP or Azure?

On AWS and Google Cloud, the docs cover ECS or Fargate with EFS or RDS, and Cloud Run with Cloud SQL. There is no Azure-specific page; the same container image and Kubernetes manifests apply to any cluster you run there.

Is every part of AgentField open source?

The control plane and the Python, TypeScript and Go SDKs in the Agent-Field/agentfield repository are Apache 2.0. You can run it locally, self-host it and build on every documented surface without a licence key.

Is there a managed version?

No. AgentField is self-hosted only today. If you want a fully managed agent runtime, cloud vendor platforms and LangSmith Agent Server are hosted options.

Sources

Checked 2026-11-21 against the AgentField docs and each project's own repository or docs on that date.