--- title: "Cielara Code Targets AI Coding Tools' Biggest Weakness: Finding the Right File To Edit" description: Causal Dynamics Lab research shows AI agents spend 80% of time searching files. Cielara Code uses causal graphs to outperform Claude Code and OpenAI Codex in benchmarks. author: Darie Nani (Editor-in-Chief) date: 2026-05-05T19:30:56.643Z updated: 2026-05-05T19:30:56.653Z canonical: https://www.sovereignmagazine.com/article/cielara-code-ai-coding-localization image: https://cdn.nanimediahouse.com/cdl-team-photo.webp categories: Artificial Intelligence, Startups content_type: Spotlight region: San Francisco publication: Sovereign Magazine about: - type: Organization name: Causal Dynamics Lab description: Causal Dynamics Lab builds validation infrastructure for AI-generated software. Its platform, Cielara, uses a graph-structured causal memory system to predict how proposed code changes will behave in production before they ship. Founded by former Uber platform engineers and AI researchers from Microsoft Research and Emory University. Headquartered in San Francisco. url: https://causaldynamics.com industry: AI Developer Tools sameAs: - https://causaldynamics.com --- AI coding tools have an efficiency problem that most users never see. According to research from Causal Dynamics Lab, a San Francisco-based startup, [AI coding agents](https://www.sovereignmagazine.com/article/potpie-ai-raises-2-2-million-to-give-ai-agents-codebase-context) spend the majority of their working time not writing code but searching for the right files to edit. The company's analysis of thousands of coding sessions found that 56.8 per cent of all agent actions were file reads and 24.2 per cent were grep (keyword text search) queries. Less than one per cent of actions were actual code edits. The agents were not struggling to generate patches. They were struggling to locate where those patches should go. Causal Dynamics Lab says its new product, Cielara Code, addresses this by giving AI coding tools a structural map of production software rather than forcing them to scan files sequentially. ## How AI Code Generation Tools Lose Accuracy on Complex Tasks The inefficiency compounds as task complexity increases. When the correct fix spans more than six files, agent recall drops from 0.579 to 0.143, according to CDL's research. Failed attempts consume four times the compute of successful ones. The 2025 DORA report found that higher [AI coding tool adoption](https://www.sovereignmagazine.com/article/anthropic-study-finds-ai-coding-assistants-speed-up-work-but-reduce-skills) correlated with a 7.2 per cent decline in deployment stability. Werner Vogels, AWS's chief technology officer, has described this dynamic as "verification debt," where development teams write less code themselves but must spend more time reviewing machine-generated output they did not author. Separate data from Sonar found that 96 per cent of developers do not fully trust AI-generated code is correct, while only 48 per cent say they always verify output before committing. ## What Cielara Code Does Differently From Existing AI Developer Tools Rather than treating codebases as flat text, Cielara Code builds what CDL calls a Production World Model: a six-layer causal graph mapping what the code does, why it was built, who owns it, how it is constrained, where it runs, and what happened at runtime. Before an agent begins exploring a codebase, Cielara constructs a Code Dependency Causal Graph that indexes four relationship types between files. The agent navigates this structure rather than scanning files one by one. The practical effect, CDL claims, is that a runtime failure can be traced back through the commit that introduced it, the developer who approved it, and the original intent behind the change. ## Benchmark Results: Cielara Code vs Claude Code and OpenAI Codex Across three independent benchmarks, Cielara Code achieved an overall localization accuracy of 0.774, compared with 0.738 for [Claude Code (Opus-4.6)](https://www.sovereignmagazine.com/article/anthropic-and-openai-release-competing-ai-coding-models-within-minutes-of-each-other) and 0.707 for OpenAI Codex (GPT-5.4). On MULocBench, a test covering 1,033 issues across 46 repositories, Cielara reached 0.752 recall@5 versus 0.727 for Claude Code. Mean task time fell from 141.84 seconds to 128.62 seconds. CDL reports 30 to 40 per cent lower compute cost per task overall. The benchmarks measured code localization, the task of identifying where in a codebase a change needs to be made, rather than the quality of the generated code itself. ## How REASONARA Handles Enterprise-Scale AI Code Review Cielara Code's underlying memory system, REASONARA, stores production context as a graph-structured causal memory holding more than 125 million tokens of effective context. A typical lookup retrieves 1,000 to 2,500 tokens rather than the 23,000 to 115,000 tokens required by full-context approaches. On independent benchmarks, REASONARA scores 94 per cent on UltraDomain, 92 per cent on LoCoMo, 73 per cent on LoCoMo-plus, and 87.4 per cent on LongMemEval. CDL says it runs five to eight times faster than Codex in high-reasoning mode. The company's roadmap targets a one-billion-token context window. ## Who Built Cielara Code and Who Uses It Causal Dynamics Lab was founded by Hasibul Haque, who led platform engineering at Uber during its rapid growth phase, and Ryan Turner, a former Uber staff engineer and maintainer of the SPIRE project within the Cloud Native Computing Foundation. Research is led by Dr Xuchao Zhang, formerly of Microsoft Research, and Dr Liang Zhao of Emory University, who has more than 200 publications and is ranked among the top two per cent of scientists by Stanford University. CDL has a formal research partnership with Emory's AI Lab. The company says 11 Fortune 100 and more than 40 Fortune 500 companies currently use Cielara Code. It positions the product as a verification layer for existing AI coding agents rather than a replacement for them. "Every coding agent today navigates by grep. That is the equivalent of a surgeon operating without imaging," said Haque. "We built Cielara Code to give agents sight: a causal model of the production environment that makes the reasoning behind every change explicit and verifiable." Phillip Miller, global chief information security officer at H&R Block, said: "Enterprises need solutions to problems they cannot solve with people alone. Cielara's technology is a generational leap towards the original promise of AI: tackling complexity around the clock with acquired knowledge, deep reasoning, and unbeatable accuracy." **About Causal Dynamics Lab** Causal Dynamics Lab builds validation infrastructure for AI-generated software. Its platform, Cielara, uses a graph-structured causal memory system to predict how proposed code changes will behave in production before they ship. Founded by former Uber platform engineers and AI researchers from Microsoft Research and Emory University. Headquartered in San Francisco. [Website](https://causaldynamics.com) ## FAQ **Q: What is code localization in AI coding tools?** Code localization is the task of identifying where in a codebase a change needs to be made. It is distinct from code generation, which is the act of writing the change itself. Research shows AI coding agents spend the majority of their time on localization rather than editing. **Q: How does Cielara Code compare to Claude Code and OpenAI Codex?** In benchmarks measuring localization accuracy, Cielara Code scored 0.774 versus 0.738 for Claude Code (Opus-4.6) and 0.707 for OpenAI Codex (GPT-5.4). It also reduced mean task time and compute costs by 30 to 40 per cent. **Q: What is verification debt in AI software development?** Verification debt is a term used by AWS CTO Werner Vogels to describe the growing gap between the speed at which AI generates code and the speed at which teams can verify that code will behave correctly in production. The 2025 DORA report linked higher AI tool adoption to a 7.2 per cent decline in deployment stability. **Q: Who uses Cielara Code?** Causal Dynamics Lab says 11 Fortune 100 and more than 40 Fortune 500 companies use Cielara Code as a verification layer for their existing AI coding agents.