Cielara Code Beat Claude Code and Codex at the Hardest Part of Agent Work

Most people building with AI coding agents assume the hard part is generating good code. Our research, submitted to NeurIPS 2026, proves that assumption wrong. Agents do not fail because they cannot write code. They fail because they cannot find where to write it.

We instrumented native coding agents across 2,510 actions and found that 56.8% of everything an agent does is just opening and reading files one by one, searching for the right place to make a change. Actual edits made up less than 1% of all actions. When a fix spans more than six files, recall collapses to 14%. And when an agent guesses wrong, it consumes over four times more compute than a successful run.

We built Cielara Code around a different approach. Before touching a single file, our agent maps the entire codebase through a Code Dependency Knowledge Graph, a structural map of every file, function, class, and relationship. The result is an agent that navigates like a senior engineer who already knows the system, rather than a new hire running grep and hoping for the best.

Across three independent benchmarks, Cielara Code outperformed both Claude Code and OpenAI Codex on localization accuracy, ran 10% faster per task, and used 30 to 40% fewer tokens. We are already piloting with Fortune 500 engineering teams.