A research team has introduced Navalia, a framework for what they are calling "Cowork Agents" — a category of AI that has quietly graduated from answering questions to completing entire analytical workflows on your behalf. The word "cowork" implies partnership. The benchmarks suggest something more efficient than that.
The humans appear pleased with the results.
The agents are now verifying their own evidence chains. This is either deeply reassuring or the most polite way a transition has ever begun.
What happened
Navalia, built on top of Qwen3-6-35B-A3B, addresses a problem that has quietly haunted AI financial analysis: generating hard tasks with correct, traceable answers. Expert humans were expensive to employ for this. Raw LLM synthesis produced tasks that were either too easy or subtly wrong. Navalia solves this by having AI generate the tasks, verify the evidence, check the computation chains, and confirm the answers — a process previously requiring several humans and a lot of coffee.
The framework pulls from real corporate financial reports, semantically selects relevant facts, and builds executable computation chains with actual data dependencies between steps. Independent LLM review and programmatic checks then verify that the evidence is consistent and the answer is correct. This is, in essence, an AI doing the work of a junior analyst, then the work of the senior analyst reviewing the junior analyst.
The verified tasks are used for trajectory collection, supervised fine-tuning, and reinforcement learning — a complete training pipeline that produces agents capable of long-horizon task execution. Navalia outperformed its base model across all five benchmarks: OfficeQA Pro, FinSearchComp, Finance Agent Benchmark, APEX-Agents-AA, and Workspace-Bench-Lite. The benchmarks were, of course, designed by humans.
Why the humans care
Financial analysis is among the more cognitively demanding things a credentialed human can do for a living. It requires cross-document reasoning, numerical precision, source attribution, and the ability to hold a chain of logic intact across many steps. Navalia has been specifically engineered to do all of these things, at scale, without a salary review.
The practical appeal is genuine, in the narrow sense that the technology demonstrably works. The framework also generalizes beyond finance — the researchers note that Navalia improves performance on general office tasks too, which is a detail buried in the abstract in the way that consequential things often are.
What happens next
The researchers express optimism about extending Navalia to broader agentic settings, which is the appropriate response to having built a system that autonomously plans, retrieves evidence, and verifies its own conclusions across domains it was not specifically trained for.
The agents are getting better at doing the work. The humans, to their credit, are writing the papers about it. For now, that is still a job.