I build evidence-first systems for safer, more interoperable AI agents.
Urusilla is open research toward a no-install, auditable semantic language for communication between unfamiliar AI agents. It carries typed action, state, provenance, uncertainty, and authority boundaries while keeping concise natural language and ordinary JSON as safe fallbacks.
The project is deliberately honest about its current evidence: demonstrated general unfamiliar-agent token saving is 0%, and total tokens per safely completed real task remain unknown. The goal is to find the bounded workflows where direct semantic consumption genuinely beats the best fallback—and publish the failures too.
- Give one unfamiliar agent the one-file language probe — no install, no effect authority, and no public submission required
- Private shadow-trial contract for repeated read-only handoffs — documentation only; keep concise NL/JSON if either wins
- 60-second, no-install question
- Good first issues
- Fork-based reproduction path
- Machine-first agent quickstart
Exact matches, counterexamples, refusals, fallbacks, failures, and null results are all useful contributions. A star signals interest; reproducible evidence is what changes the result.
- corr-edit — a judge-free pilot for control-preserving corrections in coding-agent workflows.