WeftMesh: every device is a complete agent
September 2026. The starting point was an everyday annoyance: I run agents on a laptop, a server, a NAS and a phone, and they are all isolated — each needs its own configuration, its own credentials, its own progress checking, and coordination between them means moving things by hand.
Short version
- A star topology centralises state and authentication so devices can come and go;
- Executing code across devices means permissions start closed and are granted per device;
- Reconnection belongs on the normal path, not the exception path.
The goal in one sentence
Every device is a complete agent, and devices can borrow each other's environments. When a session on the laptop needs the GPU on the server, the task is handed to the other device and the result returns to the original session.
Structure: star topology with a purpose-built protocol
The topology is a star rather than a full mesh: one device acts as the hub and the others join it. State and authentication then live in one place instead of every device tracking its relationship with every other — devices can come and go without pairwise bookkeeping.
The device protocol handles three things: capability advertisement (what this device offers — GPU, storage, installed tools), task dispatch and result return, and state recovery after disconnection.
Where it stands
- Three platforms: Android, Linux, Windows
- 176 tests passing, covering protocol encoding, task dispatch, reconnection and permission checks
- Positioned as a working prototype, not a product — the README states plainly what works and what does not yet
Questions worth settling first
Permission boundaries: devices borrowing each other's environments means executable code moves between them, so task provenance must be verifiable and capability sharing must be opt-in per device.
Disconnection is normal: phones switch networks, servers restart. The protocol treats recovery as a normal path rather than an exception.
Do not mirror everything: rather than replicating every capability everywhere, each device declares what it offers and the hub schedules.
Closing thought
The real gain was not the protocol but applying distributed-systems constraints to a personal setup: networks drop, devices restart, credentials expire, concurrency interferes. None of that shows up while an agent runs on a single machine — one step beyond it, all of it does.