coherent
collective intelligence
When emitters fall into phase, they radiate together — and the burst scales as the square of their number. Superradiant builds agents on the same principle: systems that improve one another, then emit far more than the sum of their parts.
SIA: agents that improve agents
SIA (Self-Improving AI) closes a loop. A Meta-Agent writes and improves a Target-Agent; the Target-Agent runs the task; a Feedback-Agent reads the resulting trajectory and proposes the next improvement. Each generation makes the next one better.
The signal that drives that loop is the trajectory — every prompt, context, score, and execution log. Capture it faithfully and the system compounds. Lose it and self-improvement decays.
Read the technical paperbuilt in rust for speed, safety & observability
sia_rust is a native Rust re-implementation of the SIA core — orchestrator, context and trajectory management, prompt system, run-directory format, and visualizer — reproducing the reference implementation's outputs byte-for-byte, verified by a differential-parity harness.
The deterministic scaffolding runs roughly 5.8× faster by geometric mean across nine operations — up to ~25× on prompt building. A capability allow-list and an explicit threat model treat the execution sandbox as a first-class concern, because the agent's output is itself executable code.
See the benchmarksopen research on self-improving systems
We work in the open: a faithful, auditable core; honest engineering seams instead of overclaimed results; every not-yet-evaluated component marked as future work. The lab is small and the surface is large.
Work with us