Featured in Domain Arrivals · 2026-08-12
Agentix AI Asks Whether the Agent Actually Did the Work
An evaluation harness for AI agents that checks machine end states instead of trusting an agent’s final report.
Why it surfacedThe proposed harness has seven explicit stages, from task specification and sealed starting states to transcript capture, end-state diffs, assertion runs, and replay. It is unusually candid about its limits—and equally candid that nothing has shipped yet.
Continue into the edition
Ice Does Not Count
August 12’s newborn domains include a rule-bound archive of enormous sodas, 27 family stories cut from one afternoon, a disaster-supply map with a 24-hour freshness rule, and an ocean bell that rings only when six conditions agree.
This is one of 1,000 discoveries in the 2026-08-12 Domain Arrivals issue.