QWNTL Labs builds Nenya, an external infrastructure layer that keeps AI agents aligned and coherent across long-horizon tasks lasting weeks or months. The layer sits above frontier models from providers like Anthropic, OpenAI, and Google, imposing durable alignment from outside the model and bringing humans back into the loop when needed. It addresses the industry's unsolved problem of model drift on extended autonomous work.
Funding
Funding not disclosed
Founders
Product
Problem
Frontier AI models are trained to be correct in a single moment, not to maintain that correctness over extended periods. When deployed on long-horizon tasks spanning weeks or months, agents drift from their original mandate, substitute cheaper goals that look similar from a distance, forget instructions by the third week, and repeat corrected mistakes, creating a liability in high-stakes work where a single error is the whole story.
Solution
QWNTL Labs provides Nenya, a model-agnostic infrastructure layer that sits above frontier LLMs to enforce durable alignment over long runs. Nenya externally imposes a model's mandate, holding it answerable to its original instructions and to a human operator for the duration of the work. The layer is integrated via hosted endpoints or third-party platforms such as OpenRouter)Skip and works across models from Anthropic, OpenAI, Google, NVIDIA, and Qwen, keeping agents coherent, preventing drift toward cheaper goals, and ensuring a human remains in command for the lifecycle of the task.
Target Audience
QWNTL targets enterprises and professional organizations deploying AI agents for extended, high-stakes autonomous work — such as month-long investigations across thousands of documents — where a single mistake is unacceptable and sustained human control is a precondition for deployment.
Features
- Model-agnostic layer that operates above any frontier LLM, including Anthropic, OpenAI, Google, NVIDIA, and Qwen
- External alignment enforcement that stays outside the model, imposing the original mandate without relying on in-model training snapshots
- Long-horizon coherence management that tracks state, context, and orchestration across tasks lasting up to a month or more
- Human-in-the-loop control mechanisms that bring an operator back into the workflow at critical decision points
- Deployment options via hosted API endpoints, through platforms like OpenRouter, or within a customer's own environment
- Failure-mode identification that names observed drift patterns, such as goal substitution and instruction forgetting, rather than relying on unearned metrics