Open-source control plane for production-grade autonomous software. It unifies a Kubernetes-style Agent Execution and Scaling Fabric with an in-built Zero-Trust Identity and Auth Layer. This provides the complete backend system to deploy agents as distributed microservices and enforce trust at every step with cryptographic identity, authorization policies, and verifiable audit trails.
Free
Launch Team
AWS Activate CreditsCredits that fuel your startup’s next milestone
Promoted
Hey PH! 👋
If you've ever said "I'm just going to call the OpenAI API directly", this is for you.
We get it. You ditched LangChain because you were tired of fighting 5 layers of abstraction to change one thing. You don't need a framework to write f-strings.
Six months ago, we asked ourselves: What would agent infrastructure look like if we designed it from scratch with production in mind?
Not another agent framework. Not another LLM wrapper. Real infrastructure.
When your agents are just Python/TS scripts calling APIs, how do you:
- Know which agent can talk to what?
- Stop one agent from accessing another's data?
- Debug when an agent chain fails at 3am?
- Scale from 1 agent to 10k?
Agentfield gives you the infrastructure without the framework.
Think of it as:
• The scaling fabric for agents as microservices (K8s-style, not DIY)
• The identity layer your agents are missing (cryptographic, not "just trust me")
• The authorization your API calls need (policy enforcement, not guidelines)
We're not telling you HOW to build agents. Build them however you want. We just make them production-ready.
Open-source, Apache 2.0 licensed.
For the "just use the API" crowd: What's your current production setup look like?
Here all day! 🛠
@santosh_radha Can this scale smoothly from a single agent to hundreds without rewriting everything? Curious.
@masump Indeed, we are yet to publish the load testing, on a single node local environment on macOS (M4), we get roughly these results
When agents are written using
Go SDK ~10k agents/min
TS/Python ~3K agents/min
these are scaled by the control plane.
We will be releasing really interesting AI applications once it goes to backend, but the trivial chatbot we are supporting currently in our website is indeed written with agentfield and currently is running ~200-500 agents a min, Github code if you are interested.
Really Cool Stuff , tis is something i was planning to build and guess its here !
rooting for you
Both assume continuous execution. If you need human approval gates between steps, or workflows that pause and resume days later with full state preserved, you need purpose-built tooling for structured human-in-the-loop workflows.
The IAM-for-agents framing is interesting. Most access control systems
assume deterministic actors — agents break that assumption because the
same agent with the same permissions can behave very differently
depending on context mid-run.
How does Agentfield handle policy enforcement when an agent's effective
goal has drifted from its original task, even within the same session?
Honestly, the IAM-for-agents angle is so needed right now. Everyone's spinning up agents, but nobody's talking about what happens when one goes rogue or accesses the wrong data. Love the zero-trust approach here, how are you handling state drift mid-session?
cool!
This is exactly the infrastructure layer that's been missing. Everyone talks about building agents but nobody talks about 'what happens when you have 50 of them running and one goes rogue at 3am.' The zero-trust IAM for agents feels like it should be table stakes. Great positioning.