Work2023
Deco Agent Network
Architecture shipped; the original RAG goal paused when Hermae took focus.
Challenge
Personal assistants in 2023 were mostly a single prompt with a pile of tools glued on. If you wanted two machines to share a capability, you either shared a filesystem or you invented an ad-hoc protocol nobody else could implement. We wanted a node you could run locally, extend with plugins, and talk to other nodes without handing over the whole disk.
The original product goal was larger than a framework: RAG plus inter-server LLM calls, so a personal network could retrieve and hand work across machines. That's an architecture problem first. If the contracts are wrong, retrieval never has a safe place to live.
Approach
Through late 2023 we built Deco Agent Network, a server framework for personal assistant nodes. Plugins were manifests that extended what a node could do and described safe cross-app APIs, so one assistant could request a capability from another without sharing a full filesystem.
The repo was a Turborepo with a core server, a plugin SDK, and example nodes. The design choice that mattered was the manifest, because a capability had a name and a way to refuse. We still use that habit on AI plugins: when a product needs to be invoked from Claude, Cursor, or ChatGPT, expose the job, not the account.
Outcome
The architecture shipped. The original RAG goal receded when Hermae took the calendar, and Deco still taught us to keep assistant capabilities behind explicit contracts instead of one mega-prompt.
This isn't a market win. It's here because later work kept the plugin surface. If you're designing agent systems or connectors, the part worth keeping is the contract, not the repo name.