Labs
Building a data flywheel for our agent harness
An observability pipeline that reads production traces at corpus scale, names recurring failure modes, and proves the fix held.
Aug 22, 2026 · 7 min read
September 10, 2026 4 articles
Labs
Research, experiments and benchmarks from the team building the platform.
Labs
Using semantic benchmarks to build a self-improving text-to-query agent
Our text-to-query agent went from 45 seconds on frontier models to 2 seconds on GLM 5.3 Flash, with the same accuracy at 1/20 the cost. This post describes the architecture, evaluation loop, and lessons learned.
Levi Kline September 2, 2026 10 min read
Labs (3)
Labs
What we took out of the context window
A median request carried 150,000 tokens. Most of our context work since comes down to one move, applied repeatedly.
Aug 14, 2026 · 7 min read
Labs
From an agent harness to a research platform
How and why we made our harness model-agnostic, and what per-turn portability actually cost us.
Aug 6, 2026 · 7 min read