Winter 2026
SIGIL
A shared store where AI agents claim work by tagging it, instead of talking to each other — built to measure whether that actually helps.
SIGIL — Structured Inter-agent Graph & Index Layer — asks what happens when several AI agents work on the same task at the same time. Instead of passing messages or sharing full context, agents claim ownership of work by writing explicit tags into a deterministic, tag-queried shared store.
That idea isn't new: it's a deliberate revival of the blackboard and tuplespace architectures that predate most of what the field is currently excited about. The team chose a conventional architecture on purpose, because the point of the project isn't the design.
The point is the measurement. It's an empirical study of whether explicit, deterministic ownership makes multi-agent systems more reliable and cheaper to run under contention — when agents are actively competing for the same work. The team built the coordination store, a headless agentic harness to drive it, fixed task-DAG scenarios to run against, and the metrics layer to record what happened, then presented their findings at the showcase.
What the team had to work out
- Freezing a shared contract early enough that language-agnostic components could be built against it independently
- Designing scenarios where contention is real and reproducible, not incidental
- Measuring reliability and cost in a way that survives the nondeterminism of the agents themselves
Other builds
Build the next one
Applications for each projects intake go out to members first. No experience assumed — there's a beginner track for exactly that reason.