Retries, state, auth, rate limiting, logging — built in. Stop reinventing the scaffold. Ship the agent.
from chassis import Agent, tool, state
agent = Agent("support-bot", state=state.redis("prod"))
agent.register(tool(search_kb), tool(create_ticket))
agent.run(max_retries=3, auth=jwt, rate_limit="100/min")
The Problem
Engineers building AI agents spend 80% of sprint capacity on scaffolding that has nothing to do with the agent's actual job.
Exponential backoff, jitter, dead-letter queues — written from scratch, tested inadequately, and abandoned when the real work starts.
Conversation history, tool call state, memory — bolted together with Redis, Postgres, and S3 in ways that break under load.
Custom scaffolding becomes legacy code. Security patches, framework upgrades, oncall incidents — for infrastructure that isn't your product.
What's in the chassis
Six hardened primitives. Drop in your business logic.
Configurable retry policies with exponential backoff, jitter, and dead-letter routing. Works across any LLM provider.
Conversation history and tool call state stored in Redis, Postgres, or in-memory. Survives restarts and horizontal scaling.
Structured tool registration with JSON Schema validation, version pinning, and automatic error surfacing to the model.
JWT, API key, and OAuth flows handled at the chassis layer. Secrets never touch your agent logic.
Per-agent, per-user, and global rate limits. Token budget tracking. Graceful degradation when limits are hit.
Every tool call, LLM request, retry, and error emitted as structured events. OpenTelemetry-compatible out of the box.
How it works
One package. Zero config required to start. Sane defaults for retries, state, and logging that work out of the box.
Wrap any function as a tool. The chassis handles schema generation, input validation, and error propagation automatically.
Your agent code is pure intent — what the agent should do, not how to recover when things break. That's our job.
From the blog
The gap between a working demo and a production agent is exactly six problems. Here's what they are.
Read article →Most frameworks were designed for demos. When traffic arrives, three assumptions break immediately.
Read article →What state means for agents, the four approaches, and which one survives production.
Read article →Open-source and free to self-host. Managed cloud starts at $79/mo.
Get the framework