The Eval Stack: Proving AI Agents’ Rightness
July 10, 2026 – 4:23 pm
Image by: SixtyFour
Most AI research tools rely on language models to browse the web and produce outputs. Saarth Shah, however, built Sixtyfour with a contrasting approach: grade everything and only ship what improves the score.
Saarth Shah maintains a scoreboard for every build of SixtyFour’s research agents. These agents are evaluated against questions crafted by hand by experts and checked against real-world cases across various sectors. The grading system determines which updates ship, ensuring quality over quantity.
Initially, the industry attributed issues to hallucinations—models inventing facts and sources. While this problem persists, frontier models now hallucinate less, and combining them with live web search has narrowed the gap. Yet, Shah’s primary concern is not model imagination but their reach.
A language model with a search box only accesses what a user with a browser can see. Crucially, many facts crucial for investigations lie below the surface in licensed records, unindexed filings, and complex networks that fraud attempts conceal. Surface research cannot uncover these.
"The models keep getting better, and they still only see what a person with a browser can see," Shah remarked. "The answers that truly matter are deeper, and you have to actively seek them."
Approximately 96% of enterprises now employ AI agents, yet the gap between capable systems and those that deliver accurate results on complex professional tasks remains significant. Years before benchmarks confirmed this, Shah realized: capability and reliability are distinct traits, and the latter requires deliberate engineering.
Sixtyfour offers a unique system architecture. Its agents bridge two failing options: static databases that go stale and general models that improvise. They combine live web research with public records, state and legal filings, open-source intelligence, and licensed data to produce reports where every claim is traceable to its source—a level of transparency akin to what a court might demand.
In practice, a single query triggers a cascade of investigations. If Sixtyfour is asked to verify a seller, it pulls corporate filings, cross-references addresses and officers, scrutinizes sanctions and litigation records, and searches the live web for comprehensive insights. Agents follow threads until they can conclusively answer a question rather than stopping at initial plausible answers. The output is a detailed dossier, allowing reviewers to audit the machine’s reasoning instead of blindly trusting it.