Craft HackAIthon entry: Magency, a Nostr client for AI agents
There is no good way to control AI agents on a mobile phone. There should be an app for that!
Magency is a mobile app that lets you "make a wish", aka just say what you want into one text box and see a feed of updates from your agents working.
Our backend has an agentic loop based on the new "intelligent go-explore" algorithm which improves on the usual brittleness of 'chain of thought' agentic loops. More on IGE here: https://arxiv.org/abs/2405.15143
We combine IGE with a new idea: what if agents can access a shared global archive of events - so they can learn from each other - as well as a shared global skill library?
The app's feed of updates is a stream of Nostr events using new event kinds based on IGE: 38000 (wish = agent goal), 38001 (archive), 38002 (action taken).
Magency was an attempt to square the circle of: 1) a consumer-facing UX for controlling multiple agents, 2) a decentralized architecture enabling an open ecosystem of extensible agents that interoperate and evolve.
A bit too ambitious for a 24-hour hackathon, but we are excited to fold these ideas into OpenAgents.
https://stacker.news/items/564174/r/AtlantisPleb
Demo video on X: https://x.com/AtlantisPleb/status/1798730530374230312
Christopher David /
npub1tl…k7fdm
2024-06-06 15:15:26
Author Public Key
npub1tlv67m7xvlyplzexuynmfpguvyet0sjffce3y8vu0suuyuwgzauqjk7fdmPublished at
2024-06-06 15:15:26Event JSON
{
"id": "e76a9751540edc244661651891c9039ccc433498d25650a392752ba32ab228fc",
"pubkey": "5fd9af6fc667c81f8b26e127b4851c6132b7c2494e33121d9c7c39c271c81778",
"created_at": 1717686926,
"kind": 1,
"tags": [
[
"r",
"https://arxiv.org/abs/2405.15143"
],
[
"r",
"https://stacker.news/items/564174/r/AtlantisPleb"
],
[
"r",
"https://x.com/AtlantisPleb/status/1798730530374230312"
]
],
"content": "Craft HackAIthon entry: Magency, a Nostr client for AI agents\n\nThere is no good way to control AI agents on a mobile phone. There should be an app for that!\n\nMagency is a mobile app that lets you \"make a wish\", aka just say what you want into one text box and see a feed of updates from your agents working.\n\nOur backend has an agentic loop based on the new \"intelligent go-explore\" algorithm which improves on the usual brittleness of 'chain of thought' agentic loops. More on IGE here: https://arxiv.org/abs/2405.15143\n\nWe combine IGE with a new idea: what if agents can access a shared global archive of events - so they can learn from each other - as well as a shared global skill library?\n\nThe app's feed of updates is a stream of Nostr events using new event kinds based on IGE: 38000 (wish = agent goal), 38001 (archive), 38002 (action taken).\n\nMagency was an attempt to square the circle of: 1) a consumer-facing UX for controlling multiple agents, 2) a decentralized architecture enabling an open ecosystem of extensible agents that interoperate and evolve.\n\nA bit too ambitious for a 24-hour hackathon, but we are excited to fold these ideas into OpenAgents.\n\nhttps://stacker.news/items/564174/r/AtlantisPleb\n\nDemo video on X: https://x.com/AtlantisPleb/status/1798730530374230312",
"sig": "7166558afd4da54771b7a2d441468f4673de2275aa197b4459dd78e518a8c5ed8de257beb32264da5607e71eb4455ff406c89a2dd952e2f8da9e8f6dd20de7bb"
}