agentic-kitA durable runtime for AI agent runs

Not an agent framework. It owns the lifecycle of a run — that it outlives the request that started it, survives a worker dying mid-step, can be stopped, parked for a human, resumed exactly where it stopped, nested, metered, and watched by several people at once.

Get started

Core concepts

GitHub

Runs outlive the request

run() persists, enqueues and returns in milliseconds. The model call happens somewhere else, later, in a worker that can die and resume from the last committed step.

A human in the middle

A parked approval is a durable state holding no process — no worker, no lock, no memory. It expires on a queue timer whether anyone is watching or not.

Subagents are runs

Same loop, same table, same persistence. So a subagent can use tools, park for a human, and be resumed — none of it built twice.

Your database

Four small interfaces stand between the engine and your stack. Postgres, Mongo, Dynamo, SQLite — the engine imports no driver.

Every client agrees

Hydrate from the durable log, tail the bus from a cursor. Two tabs, a reload, and a mid-run reconnect all converge on the same conversation.

Bring your own UI

use-agentenkit is a React hook with every route, label and transport under your control — and the follow's frames are a public contract you can build any client over, in our shape or as AG-UI.