Enterprise Environments

The core problem with enterprise agents is not simply that they lack intelligence, but that they fail to understand and operate within enterprise workflows. An enterprise workflow typically involves a living system of policies, hierarchies, conventions, and dependencies that dictate how tasks get done. Knowing these workflows is essential, because an agent cannot simply brute-force its way through sprawling enterprise data. It must know where to look, how documents are organized, and when to pivot to different sources. And crucially, this knowledge can't be front-loaded in one giant context dump. In practice, the agent must pull in workflow knowledge at runtime as necessary, just as an experienced employee would recall the right policy or tool only when the situation demands it.

Raising the Bandwidth of Learning

This is why environments are necessary for agents to learn, and why the bottleneck is the learning loop we run inside those environments. Reinforcement learning became the favored tool because it compresses supervision into a single scalar signal—a verifier function that rewards success. But in enterprise settings, this is a mismatch. A single score at the end of an hour-long rollout is an extraordinarily low-bandwidth signal given the richness of what happened along the way. Humans, when completing complex office tasks, don’t simply mark them as “done” or “not done.” They consider what worked, what failed, and what patterns might apply again later. These lessons accumulate as reusable instructions, tied to recognizable states. Over time, with repetition, they distill into intuition, the way a financial analyst no longer consciously recalls each rule when building a model but nonetheless follows them. The goal is not to abandon RL but to raise the bandwidth of what we learn per rollout and to better align it with the human pattern of improvement—one where partial lessons carry forward, not just final outcomes.

Environments that Mirror Enterprise Reality

Aviro’s Cortex is the first attempt at doing this. We built environments specifically tuned to enterprise. We constructed deep search environments where the puzzles require mastering enterprise workflows, not just surface-level retrieval. To do this, we created three hypothetical firms—in finance, consulting, and accounting—complete with realistic enterprise data, built with the input of domain experts. The environments deliberately replicate the documentation complexity of large organizations. Policies are siloed, systems are fragmented, and the relevant answer is often several hops away across tools. We embedded workflows such as: “legal policies for X are always filed in Y,” or “sales performance data must be cross-checked against approved spreadsheets.” These rules reflect how real organizations codify their processes, and an agent that doesn’t internalize them fails.

To test agents, we built environments that mirrored the systems enterprises actually use:

Email systems (Outlook, Gmail), seeded with corpora like Enron emails.
Legal policies spread across disconnected repositories.
Code documentation inspired by developer resources such as Supabase’s docs.
Operational runbooks that dictate resolution paths.
Spreadsheets (Excel, Google Sheets) for financial or reporting data.
Documentation platforms (Confluence, Google Drive, Box, Notion) housing overlapping, sometimes conflicting references.
Ticketing systems (Jira, Linear) that encode workflow state.

These environments force agents to plow through multi-hop, multi-system reasoning challenges that define enterprise knowledge work. The first experiments used a deep research agent that had to operate under these conditions. The criteria for success required the agent to accurately determine where in the enterprise corpus to search next, when to escalate an issue to a human or another system, how to cross-reference information from disparate sources, and how to adapt its strategy when encountering ambiguous or incomplete data. Enterprise learning doesn’t come from static instructions alone, rather also emerges from runtime adaptation inside environments that faithfully reproduce the messiness of organizational systems.

By embedding workflow knowledge into environments and capturing lessons as reusable state-keyed instructions, Aviro’s Cortex closes the gap between reinforcement learning’s low-bandwidth signals and the high-bandwidth, context-sensitive way humans actually improve. In enterprises, where workflows are the backbone of every task, this shift is the only way to make agents learn and adapt in a way that scales.