An observatory for heterogeneous agent teams. You approved the run — the swarm is yours now. Every artifact below traces to a real run in the repo: a 4-agent swarm audit, cross-vendor review councils, caveman compression A/Bs (the playground is a live JS approximation of the rules — everything else is the genuine output).
The loop, explorable
dispatch → review → crosscheck → gate, with the real payloads from today's runs — including the words-vs-tokens catch.
new · playgroundcaveman compressor
Paste prose, watch it shed words the way K3 does it: −46% condensing, +11% generating. Genre decides, live.
new · boardswarm audit
3 P0 · 8 P1, found by 4 read-only K3 subagents. 6/6 citations verified. Open the board.
landingSignal / Terminal / Paper
Three zero-animation design directions for the agentic-loops story — the restrained ancestors of this page.
capabilityKami3 / K3 hub
Long-context review, AgentSwarm exploration, quota-aware fallback, prompt-to-site — one page per pattern.
docslab notes
What the demos are, how they're deployed, and the safety boundaries they keep.