Will Brown on Training Models That Learn From Organizational Work

I lead applied research at Prime Intellect. At this conference, people compare what works, what does not and what comes next. I want to look at that question at the organizational level: what would it mean for an established enterprise to become truly AI-native?

We have moved from chatbots through reasoning systems toward agents and tools that can carry out more of a workflow. The next question is how agents can be integrated throughout an organization, not just used by a small startup building quickly.

Prime Intellect builds infrastructure for training, deploying and improving models, particularly open models and tools. Our premise is that an organization can develop models around its own workflows, data and harnesses, alongside the frontier models it already uses.

An organization contains more knowledge than its formal documentation. Some lives in code, email and written processes; some is implicit in meeting notes, relationships, local preferences and an informal understanding of who actually makes a decision.

Individual employees may see only part of an inefficient process. An agent that examines unstructured material across teams could help surface repeated friction or reveal that one group has already solved a problem another group is still facing.

I do not want this to be a one-time analysis. The goal is a continuing loop in which agents receive the context they need and both the agents and the organization improve. Human coworkers learn on the job; current agents mostly do so through comparatively limited in-context methods.

An agent can read a skill file or inspect a codebase, but it may repeatedly rediscover institutional knowledge. Each run can require processing substantial context again. Some recurring patterns, especially preferences that are difficult to state as rigid rules, may be candidates for model training.

I think of a spectrum from rules to patterns and from code to data. A skill or explicit rule belongs toward one end; preferences, friction, heuristics and trends are often distributed across data that employees learn from experience.

That raises a practical question: what belongs in context and what deserves to be learned in the weights? Not everything should be trained in immediately. But training is already a major reason new model generations improve, and organizations can consider it for repeated, valuable workflows.

Consider how a frontier model might learn a specialized enterprise workflow today. An expert shares knowledge with an expert network, an environment vendor turns it into a training environment, and a lab researcher incorporates that environment into a later training run.

That long chain can resemble a game of telephone: the person who knows the work is not necessarily supervising how the model learns it. I see an opportunity to shorten the loop between a domain expert’s feedback and durable model improvement.

With coding agents, we already correct repeated mistakes and express preferences in context. Some guidance belongs in an agent instruction file, but those files can accumulate and need maintenance. A training pipeline could also learn from examples of what a user preferred in a specific situation.

There are several complementary places for knowledge: skills and memory in the harness, documents and databases retrieved with tools, and the model’s own weights. Each has different properties; a model’s training cutoff, for example, can force it to reconstruct recent context repeatedly.

For organizations that want to train on their own work, evaluations are a useful starting point. Think of them as test cases: a model can have a lifecycle in which an organization adds examples, checks behavior and improves the system over time.

Using agents produces traces containing prompts, actions and feedback. Those traces can become evaluation assets, although curating good evaluations has historically taken work. Agents may help prepare examples, with attention to robustness and whether the results can support learning.

Decision models and routing provide a concrete example. A large and small model can attempt the same task, and feedback can identify cases where the smaller model performs well enough. Those comparisons can inform future routing decisions and cost trade-offs.

A decision model could also learn from an organization’s approvals and preferences. The relevant decisions are not limited to tool security: each organization has its own rules about its work. Feedback must be meaningful in that specific context.

For more ambitious training, production traces from a high-volume workflow can help build richer environments combining real and synthetic material. With suitable orchestration, tools and human involvement, agents can help turn those traces into environments for longer-running tasks.

I am also interested in interfaces that make human decisions easier to give. Instead of searching through long agent threads for a question, an agent might generate a small, task-specific card or widget and return when it needs a person’s judgment.

We keep moving between a single agent stream and many parallel agents. The underlying need is to give agents ideas, let them work and receive decisions at useful moments. A clear interface can make that interaction more manageable.

Those decisions can also provide explicit feedback for training a decision or orchestration model. This is not a complete solution to reward hacking or other problems, but it offers a concrete place to begin collecting labels aligned with what a person chose.

I mention Paper Instruments, a company that used some of our tooling and reported training a knowledge-work model for about $300,000. The example suggests an investment on the scale of a hire, though the right economics depend on the organization and task.

Our prototype for continual learning starts by collecting traces, flagging issues and proposing experiments. A team could review an experiment and start an overnight training run, potentially using capacity that is less busy serving employees at night.

The sequence I envision is to understand the workflow, create skills, build evaluations and routing, and then train where there is enough repeated data and value. The aim is for institutional knowledge to compound into better agents over time. That is the direction we are working toward, not a claim that every organization already has it.