Companies have spent half a century learning to manage two kinds of data.
Structured data came first. Customers, transactions, inventory, claims, and accounts became rows in databases. Once the state of a business was structured, it could be queried, measured, and turned into software.
Most of what a company knew remained elsewhere. It lived in documents, conversations, images, video, research, and code. AI is now making that unstructured material usable. Models can parse it, search it, combine it, and retrieve the relevant parts as context.
Agents produce a third kind of data: agentic exhaust.
Every goal leaves a trace: a record of what the agent searched, tried, changed, and observed. Read over time, the trace reveals a trajectory—the path the agent took through the work. Across many agents and goals, those trajectories form a cloud of agentic exhaust.
The exhaust becomes valuable when a company can search it, find a lesson, and change the next attempt.
Life sciences has accumulated a version of this exhaust for decades.
The public sees the drug that survived. Inside the company sits the history of everything that did not: targets abandoned, candidates killed, doses that failed, adverse events that changed the program, patient subgroups that responded differently, and experiments that invalidated a promising theory.
A competitor can read the approved label and published trials. It cannot reconstruct the years of negative evidence and expert judgment that shaped the final program. Reproducing that history would require capital, patient participation, and time that cannot be recovered.
Much of it never becomes public. The FDA recently found that 29.6 percent of studies highly likely to face mandatory reporting requirements had no results submitted to ClinicalTrials.gov. The missing results distort the apparent record by making successes easier to see than failures.
The private record is valuable because a failed trial can change what happens next. It can eliminate a target, redirect molecule design, alter a dose, reveal a biomarker, narrow a patient population, or force a different endpoint.
But the failure does not explain itself. The target may have been wrong. The molecule may not have engaged it. The dose may have been insufficient. The trial may have selected the wrong patients or measured the wrong outcome. Finding the lesson requires causal judgment.
Agentic exhaust creates the same problem at machine scale.
One completed task may sit at the end of hundreds of searches, tool calls, failed actions, and changes of direction. A University of Washington study of 4,300 coding-agent sessions captured roughly 350,000 model steps and 430,000 tool calls. The researchers used those traces to identify opportunities in context management, caching, tool latency, and inference serving that could not be inferred from the completed code.
As agents run longer and cross more systems, the exhaust will grow much faster than the artifacts they produce.
Most of it will be worthless. Routine successes teach little. Temporary outages create lessons that expire quickly. Repeated context and meaningless retries add volume without adding knowledge.
The valuable signal lies in the rare failure that exposes a missing capability, the correction that contains undocumented expertise, the source that changes the decision, and the recovery strategy that works across models.
Five things turn that exhaust into durable advantage.
1. See the trajectory
A useful trace begins with the objective and continues through the outcome. It preserves the context available at each decision, the paths explored, the tools selected, the state changes made, and the points where a person or another agent intervened.
No participant automatically sees the whole thing. A model provider sees its calls. An application sees the workflow. An observability system sees execution. The customer sees the downstream consequence.
The strongest position belongs to the system that can connect them.
2. Know the outcome
Completion is a weak label. An agent can finish a task while introducing a security flaw, misleading a customer, or producing work that someone quietly replaces a week later.
Useful traces connect behavior to authoritative results. The code remained in production. The customer renewed. The claim survived review. The transaction settled without fraud. The patient improved.
Without the outcome, the trace records activity. It cannot tell the company which behavior deserves to be repeated.
3. Find the durable lesson
Similar failures can come from bad context, missing permissions, unreliable tools, brittle workflows, or absent model capabilities. Each requires a different response.
The useful patterns are the ones that recur or carry unusual consequence. A tool is consistently called too late. A human correction exposes a rule nobody documented. A rare failure predicts an expensive downstream problem. A recovery strategy continues working after the base model changes.
The company must extract a lesson, test it against other traces, and reject changes that merely overfit yesterday’s failure.
Two measures matter. Lesson yield is the share of traces that produce a validated, reusable improvement. Write-back latency is the time between observing a failure and changing production behavior.
A company with billions of traces and no validated changes has accumulated storage, not capability.
4. Own the right to learn
Agentic exhaust can contain source code, credentials, customer records, internal communications, medical decisions, and the reasoning behind consequential approvals.
Permission to process this information does not settle who owns the learning derived from it.
Who owns an eval created from a customer failure? Can a vendor train a shared model on an expert correction? Are customer-specific rules, memories, and adapters exportable? If the underlying trace is deleted, must the derived learning also be removed?
These questions will become procurement terms because they decide where the advantage compounds.
A company that owns the traces, evals, derived rules, and adapted systems can change vendors without abandoning its learning history. If the vendor owns them, the customer may have supplied the work and judgment while the vendor accumulated the capability.
5. Change the right layer
A lesson has value only when it changes the next attempt.
If agents repeatedly struggle with an API, permission flow, or ambiguous error, repair the environment.
If they receive the wrong information, change retrieval, memory, compression, or source authority. Researchers have already trained better retrievers from agent trajectories, using browsing decisions, rejected results, and post-browse behavior as the successor to human click logs.
If the model has the capability but the workflow fails, change the harness. Adjust the instructions, tool descriptions, routing, verification, budgets, or division of work between agents.
If the agent repeatedly reaches a boundary it cannot cross, build a tool, skill, protocol, or simulator that gives it another way to sense or act.
If the capability belongs inside the model, turn successful trajectories into supervised fine-tuning data, corrections into preference data, or repeated workflows into reinforcement-learning environments.
The system has to decide where each lesson belongs. Sending every failure into model training would be slow, expensive, and often wrong.
A recent trace-guided harness repair study used failed trajectories to identify the responsible steps, patch the relevant harness layer, and reject changes that regressed on held-out tasks. Across four benchmarks, the authors report improvements of 15.2 to 50 percent over the starting harnesses.
Together, these five conditions create the monopoly.
The system that performs more consequential work encounters more edge cases, receives more expert corrections, and observes more outcomes. The lessons extracted from those traces improve the system. Better performance earns it more work, producing a richer set of traces.
A competitor can rent the same base model and copy the visible product. Synthetic data can reproduce common cases. Neither can recover the private sequence of failures, interventions, and outcomes that taught the incumbent how a particular environment behaves.
The monopoly is local. A company does not need to own all trace data. It needs exclusive access to the learning signal inside a valuable workflow.
This becomes existential when intelligence is available to everyone. If every competitor can call the same models, access to intelligence stops differentiating the company. The durable difference is how quickly each company can turn private experience into better behavior.
More business performance will move into models, context systems, harnesses, tools, evals, and policies. All of them can change in response to experience.
Every company therefore needs to become a lab. It may never train a frontier model, but it must know how to instrument its work, preserve the useful traces, connect them to outcomes, and maintain the systems that learn from them.
If an outside system performs the work, retains the traces, observes the outcome, and deploys the lesson, the vendor accumulates the capability. The company can keep its databases, documents, and finished artifacts while losing the process by which it becomes better at what it does.
Every company becomes a lab, or becomes training data for one.