The archive · AI & Models · Product decision · 2023
Prefect open-sources Marvin, betting 'AI functions' make the LLM a runtime
Prefect open-sourced Marvin, the framework behind its support Slackbot, betting 'AI functions' and scoped GPT bots become how teams build AI features.
Prefect
What the business is
Prefect builds open-source Python tooling that coordinates and observes complex data pipelines; Marvin, released under PrefectHQ, is its framework for producing structured LLM output and building deployable GPT-4/GPT-3.5 bots, born from the Slackbot that automated Prefect's own support.
How it started
Marvin began as Prefect's internal framework: it powered the Slackbot that automated support for an open-source community of close to 30,000 members. As Prefect's customers started building AI stacks, the company kept hitting the same set of problems it had solved for data engineering — integrating brittle, non-deterministic LLM APIs into well-structured pipelines. The 'lightbulb moment' was AI functions: functions with no source code that use the LLM as a runtime, which Prefect open-sourced with that idea as the headline.
What happened
By 2023-04-08, weeks after the announcement post, Prefect shipped 'marvin chat', a full-featured TUI built on Textual for chatting with GPT-4 or GPT-3.5-powered Marvin bots. The Show HN thread framed bots as mini 'conversational applications' with distinct personalities, instructions and plugins, and the TUI as a first-class way to explore and assess them. In the thread jlowin said open-sourcing had drawn an 'amazing' response, that the diversity of customer objectives gave Prefect confidence, and that the roadmap was Marvin's core — bots, plugins and knowledge handling — plus threads supporting multiple bots and users.
How it ended up
The source records no exit or shutdown: as of 2023-04-08 Marvin was a live, open-source Prefect product, with the TUI milestone shipped and the roadmap pointed at bots, plugins and knowledge handling.
Background
Marvin started as Prefect's internal framework — the code behind the Slackbot that automated support for an open-source community of close to 30,000 members. When Prefect's customers began building AI stacks, the company recognized the same problem it had spent years solving for data engineering: how to integrate brittle, non-deterministic, externally stateful APIs into well-structured pipelines.
The answer Prefect open-sourced was 'AI functions' — functions with no source code that use the LLM as a runtime. The Show HN framing by jlowin was deliberately simple: at its core Marvin is a library for building and deploying bots, and AI functions are just one kind of bot. The goal was to codify the high-level patterns Prefect kept redeploying, from structured outputs to conversational agents for B2B customers.
Weeks after the announcement, on 2023-04-08, Prefect shipped 'marvin chat', a full-featured TUI built on Textual for talking to Marvin's bots, powered by GPT-4 or GPT-3.5. The thread positioned each bot as a mini conversational application with its own personality, instructions and plugins, and the TUI as the first-class way to explore them. Prefect said the response to open-sourcing had been 'amazing' and that its next focus was the core of Marvin — bots, plugins and knowledge handling — with threads designed to host multiple bots and users.
The case is a live one: as of the thread date Marvin remained an open-source Prefect product with no recorded shutdown, and the bet — that AI stacks would need the same coordination patterns data stacks did, delivered as high-level Python primitives — was being tested in public.
What has to be true
- The origin was internal and real: Marvin already ran Prefect's support Slackbot for close to 30,000 community members, proven in production before it was ever packaged as a product.
- Customer pull preceded the pivot: it was the diversity of customer objectives — not a hunch — that made Prefect treat high-level AI tooling as a platform rather than an internal convenience.
- The bet transferred Prefect's philosophy: coordination patterns built for data stacks apply to LLM stacks, with AI functions as the small, demo-friendly primitive.
- Open source was the distribution channel: releasing under PrefectHQ with AI functions as the headline turned an internal framework into a community asset that drew the 'amazing' response.
What can be applied
A next product can hide inside your own operations: Prefect found Marvin in its support Slackbot and bet the coordination philosophy behind it transfers from data stacks to AI stacks.
Aftermath
As of the 2023-04-08 thread, Marvin was live as Prefect's open-source AI framework: the TUI had shipped, the announcement had gone out weeks earlier, and the roadmap was Marvin's core — bots, plugins and knowledge handling — plus multi-bot, multi-user threads for B2B use. Prefect positioned Marvin as a companion to its orchestration business, sharing the same Pythonic design philosophy, not as a separate company. The source records no user counts for Marvin beyond the community context and no funding or revenue figures.
Sources
spotted an error? The archive wants to know.
Your turn
You just read one. Describe what you are building, and see who is betting on the same thing.
Free account · 3 free questions · no card