Infrastructure for AI agents
Declare an agent in YAML, version it like code, and run it in production.
The problem
Building an agent is easy. Running it in production isn't.
You rebuild memory, cost, versioning, and audit for every agent. There's no Kubernetes for AI agents.
How it works
The same declarative model as Kubernetes.
Kubernetes manages the lifecycle of a container from a manifest. Vesper manages the lifecycle of an agent from a manifest.
The architecture
Seven engines. A layered platform, not a script.
Four engines ship in V1 and run an agent safely in production. Three more turn single agents into governed, multi-agent systems.
Agent Registry Engine
Declarative YAML manifests with schema validation and git-like versioning, stored in a local registry. Apply, list, inspect, and roll through versions of every agent.
Python SDK
The same agent, from Python.
Load a deployed agent and run it in three lines, or define your own tools inline. Everything the CLI does, the SDK does too.
import vesperagent = vesper.load("researcher")result = agent.run("What is the price of AAPL?")print(result.content)print(result.cost, result.prompt_tokens, result.completion_tokens)Multi-provider
Bring your own model.
Vesper routes to OpenAI, Anthropic, and Google by model name. Switch providers with a single line of YAML.