Hey HN, I've been building Sectr, a backend-as-a-service for making it easy(ier) to develop and deploy AI agents. Deploying agents requires non-trivial systems design work. Consider a support agent: how do you stream messages to your frontend, how/where do you persist messages for frontend fetching later (or for auditing), and when the agent wants to issue a refund, do you spin-down the compute infra while it's waiting on human approval? Sectr aims to be the serverless abstraction and platform to provide these systems for you.

No sign-up needed to try using the dev server:

  curl -fsSL https://sectr.dev/install.sh | sh # Install CLI
  sectr init my-agent && cd my-agent # Starter code
  sectr dev # Dev server, with UI to test chatting

Sectr provides:

- Streaming of every token, tool call and event, and persistence so they can be loaded fresh later.

- Scaling up, and down to zero when the agent needs to wait, and resumption right to the suspended agent's state.

- Human approved tools automatically use Sectr's suspension and resumption capabilities, no code needed from you.

- AG-UI endpoint for any deployed agent so you can trivially hook up your frontend

- Local dev server experience, so you can rapidly debug and iterate on your agent's behavior

At the core, you can deploy any container image to Sectr as long as it implements a certain shaped HTTP endpoint. In practice, currently Sectr provides adapters to let you deploy your existing OpenAI Agents SDK, LangGraph, or Pydantic AI agents, with only 2 new lines of code:

  app = AgentApp(name="support")
  app.add_route("/chat", openai_agents(agent))

Certainly the hardest part of building Sectr was getting the resumption logic to be correct and invisible, across the Program State x Model Provider x Agent Framework matrix. For example, an earlier bug was when a model replied with text plus a tool call, Sectr would resume that as two separate messages, instead of one message with a text and tool part. This works, but breaks prompt caching (and might subtly change model behavior).

Sectr is free during the beta (and can be tried locally without sign-up); I mainly want feedback, especially on dev & deployment experiences, any bugs, and what features would excite you.

Introductory video: https://www.youtube.com/watch?v=H-y3wynkKc0

Docs: https://docs.sectr.dev