WIPAIAI Platform
Lucia
Open-Source Builder for Long-Running AI Agents
An open-source platform for AI agents that work on one goal for weeks, follow up over email, Slack and phone, and pull in a human only when they're stuck.
The problem
A lot of work looks like "chase this until it's done": get a document back, check in with someone every week, escalate when nobody replies. Building a new agent for each of these is slow, and most agent frameworks assume a task finishes in one sitting.
The approach
- Agents are config files that get validated and compiled before they run. A new workflow is a new config, not a new service.
- A run lives for weeks and moves forward in Episodes. A scheduled follow-up, an inbound email, a call report or a human's answer each becomes one Episode, queued and run one at a time.
- Leases with fencing tokens make sure only one worker advances a run at a time. Tool calls pass through policy checks, guardrails and an idempotent outbox.
- A simulator plays the agent against LLM personas on a virtual clock, so a six-week run finishes in about two minutes. A scorecard gates publishing.
Impact
Open source on GitHub
Six weeks of simulated follow-ups run in about two minutes
One runtime runs every agent; a new use case is a config change
What I learned
- For agents that act on someone's behalf, consistency beats availability. When the runtime isn't sure of the state, it refuses to act.
- Testing an agent whose work spans weeks needs simulated time. Waiting for real follow-ups isn't an option.
Tech
PythonFastAPIPostgreSQLCeleryRedisOpenAI Agents SDKReactTypeScript