Pattern: Coding-Agent Evals Start Grading Token Discipline Alongside Correctness
TEAQL Agent Kit (2,801 stars, pushed 2026-07-27) is explicitly an evaluation environment that scores coding agents on two axes rather than one: software-engineering discipline (preserving business semantics, following generated API contracts, respecting framework boundaries, keeping audit traces, recovering from errors without unsafe shortcuts) and token-efficiency discipline — how few tokens the agent burns guessing, re-reading, or exploring irrelevant source. Its mechanism is to have the repository itself supply generated local contracts, object-specific assist output, and machine-readable evaluation reports at the moment the agent needs them. The repo also ships an AGENTS.md that instructs agents to ignore the human-facing README entirely, a small but sharp convention for repos read by both audiences.
Source
↳ Follow the thread