Save the trace - Why AI agent trajectories are your most valuable assets

Every company I talk to is racing to deploy agents, yet almost none of them are retaining the most valuable thing those agents actually produce.

Ask a team what their agent "output" is, and they'll point you to the end result, whether that's the PR that got merged, the ticket that got resolved, or the email that went out. But the agent’s work is not just the “thing” it made, but the trajectory itself. Every step it took, what it read, which tools it used, the path it chose, the false moves it made, and the reasoning it used to make all its decisions.

The trajectory itself is a valuable asset, and right now, most teams aren't saving it.

We have done this before

What’s odd is that this is not a new lesson.

Any IT administrator will tell you how important logs are. Ask anyone who has had to fix a system outage before, and they will tell you the log itself tells them what to do. The same goes for the sales reps who actively manage and review their pipelines regularly, constantly monitoring the funnel for insights rather than just assessing whether they closed the deal. Or the marketing person who is always thinking about their customer journey and is being pitched a thousand tools to monitor their conversion metrics. 

Across every org, you will see that the steps in between are just as important (and sometimes more important) as the output. Agents are the next iteration of this pattern, and their traces are just as important as the end goal.

What's different this time is what these traces actually include. A server log tells you what happened, a sales pipeline tells you what deals closed, and a marketing funnel tells you who converted. But an agent trace actually gives you the “why” of a decision. We have never had a queryable record of how knowledge work actually gets done, at each step, and at scale. We generate these records every time an agent runs, yet the default behavior is to delete them.

What the trace is actually good for

Once you start keeping trajectories, there are a million current and future use cases for them.

The most obvious one is monitoring. It is impossible to tell what an agent is doing from just the final answer alone. Generating an agent evaluation based solely on the result is equivalent to receiving a message that “the server is down” or “the customer did not make the purchase.” Without knowing the “why” behind these results, it’s hard to improve future results. And given that traces are a complete breakdown of these “whys” with way less information loss than human workflows, they are more important than ever.

Security is another important use case. Prompt injection, data exfiltration, and an agent quietly calling a tool it shouldn't; none of these flags will show up when you only view the final result. It shows up in the trajectory. If you're not retaining and constantly analyzing traces, you are letting your agents run with minimal visibility. You wouldn’t let your employee run any software on their laptop, so why should you let your agent do whatever they want without continuous oversight?

Then comes governance and audit, which should be a huge priority for teams, whether they like it or not. The first time a regulator, a customer, or your own legal team asks, "Why did the agent do that?”, the only acceptable answer is a replayable record. "We don't keep those records" is going to age as badly as "we don’t have logs."  No serious enterprise business is going to buy your product unless you can systematically show why your agent made a decision, regardless of whether the output was right or wrong.

Traces can also help you with better employee training. Every company seems to have that one employee who just “gets it” and is building faster than the others using AI. These employees aren’t using agents just for simple tasks, but are instead building out full agent factories that produce high-quality outputs. You should have these employees show the rest of your company how they are using agents through their traces. Just like with every new technology, agent utilization needs to be taught to get the most out of it. Remember, we all had to learn how to use personal computers when they were new.

The one resource that is not talked about nearly enough is internal training data. Your traces are proprietary, in-domain demonstrations of work being done in your business. This is exactly the work that is required to make Reinforcement Learning (RL), Supervised Fine-Tuning (SFT), and fine-tuning actually feasible on a per-company or per-product basis. Despite the innovation and excitement, most businesses today still use the same generic frontier models as they build their own capabilities to continuously collect and use their traces. In fact, most of the FDE-style companies that help enterprises get agents into production are effectively inspecting traces on behalf of their customers to create a feedback loop. Delete these traces, and you are just destroying your ability to differentiate over the long term.

If you still doubt that trajectories are really valuable, just consider the fact that the frontier AI labs are now paying billions for them. An entire category of companies now exists that turn traces into training environments with verifiers and then sell them. Every day, a new company starts building verifiers across every domain imaginable, and monitoring agent traces to see how they respond over time in a set environment. 

The days of humans being given tasks like “highlight the cat in this photo” or “analyze this document for mistakes” are over. Instead, we need a strong focus on the traces themselves as the valuable unit of data sold upstream. 

Software development, computer use, scientific discovery, and ML engineering are all being upended as companies build verifiers across tasks and inspect agent trajectories to assess their results. Billions are being spent today by large AI labs on these traces, further proving their importance.

Additionally, as agents go from working seconds to minutes to days, they will need all the standard infrastructure we expect from software, like disaster recovery, backups, stops and resumes, and sharing. It will be unacceptable to have an agent running for days without save states or reproducibility, and all of these must be tied not just to a single provider (the model or the client) but also to an independent party. I would argue that the trajectory itself is the agent, and that it should be saved and used however you see fit. You can start an agent on Anthropic’s models, then switch to OpenAI when a new eval drops, or use Cognition as your client, then switch to Cursor once they launch a new feature. The trace itself should be in the standard format, while the tooling around it is interchangeable. 

The real reason the trace is portable intelligence

Right now, we are in a constant state of flux with new product drops, and that doesn't seem to be slowing down any time soon. The models, frameworks, and the harness will keep changing, but the trace is the one thing that will remain constant. They are not tied to a model version or vendor, but instead serve as a record of how work gets done in your specific context.

So the real question is not "Which agent platform do we pick?" or “Which model is the best at this public eval today?” The more important question is whether you're accumulating assets while using them, and whether you're using them to the best of their ability.

What you should start doing today

Consider traces as a first-class data asset, and treat them with the same respect as your system logs, CRM, or data warehouse. Capture full trajectories by default, make them queryable, and decide your retention rules deliberately based on the agent type (instead of deleting them by default).

The companies that get this right will be way ahead of their peers and will build an asset with continuous compounding effects. This improves monitoring, security, governance, employee training, continuous improvement, disaster recovery, backups, removing lock-in, sharing, and so much more. 

You might still be early in your AI agent adoption, but whatever you do, start saving and securing your agent traces today.