I built Trytet to fix the state and geography constraints of autonomous agents.
Giving an LLM unverified host execution breaks security, and streaming embeddings back and forth over HTTP is heavy. Furthermore, when your agent hits an API rate limit or context boundary, you typically drop its execution thread entirely.
Trytet is an embeddable, sub-millisecond Wasm substrate. What it handles:
Zero-trust primitives: Evaluates volatile machine-generated code instantly without the initialization latency of Docker. Determinism: Captures active linear memory into a .tet binary. You can snapshot, hibernate, or branch an agent's execution exactly at the instruction where it halted. P2P swarming: Serializing the engine's state allows you to migrate an active agent directly to the edge node hosting your vector stores, eliminating round-trip latency manually.
The Context Router juuuust merged. It's an O(N) sliding-window estimator that enforces mathematical bounds against LLM context crashes without raw memory overhead.
Repo: https://github.com/bneb/trytet Demo / Architecture: https://trytet.com
Happy to answer questions on the engine internals or the LSM-Vector VFS hybrid.
P.S. This is active development and I'll be back tomorrow with something cool.