witanlabs/research-log

Leverages a multi-agent architecture with a REPL-based execution layer instead of discrete tool calls—allowing agents to write and execute inline scripts that compose operations more efficiently than sequential API requests. Implements structured reasoning workflows (disambiguate → define end state → plan → execute → verify) paired with domain-specific financial knowledge as reusable prompt skills, which proved more impactful than tool improvements alone. Uses programmatic benchmarking with pixel/layout/sequence comparison rather than LLM-as-judge evaluation to reliably detect regressions and guide development—the evaluation framework itself became critical to discovering architectural breakthroughs (74%→92% accuracy).