Skip to content
Outrich Data

Selling AI agent logs and trajectories

Updated 2026-11-17 · 8 min read · By the Outrich Data team

An agent trajectory is the step-by-step record of an AI agent doing a task: prompts, tool calls, results and the final outcome. Some buyers now state that they license this data from companies. The logs often hold customer data and credentials, so rights review and cleaning come first. This is general information, not legal advice.

This page is general information, not legal advice. Whether your company can license a given dataset depends on your contracts, your privacy notices and the law where you and your customers are. Review it with your own counsel before signing anything.

What agent trajectory data is

An AI agent is a model that takes actions in software to finish a task. It reads an instruction, decides on a step, calls a tool, reads the result and repeats until the task is done. The full record of one run is called a trajectory or a trace. Our glossary has a short definition of an agent trajectory.

Companies that run agents in production produce these records every day. A support agent that looks up an order and drafts a reply leaves a trace. So does a coding agent that reads a repository, edits files and runs tests. SignalFire, a venture firm, described a trace on July 13, 2026 as every step the agent took, what it read, which tools it used and the reasoning behind its decisions.

Three kinds of record are usually grouped under this heading. Agent logs are the raw traces written by the agent framework or an observability tool. Tool-call traces are the structured requests and responses between the agent and the systems it used. Evaluation transcripts are runs that have been scored, by a person or by an automated check, with the score attached.

Trajectories differ from the workflow logs a business system writes. A workflow log records what people and software did in a system of record. A trajectory records what a model did, including the text it produced along the way. Troveo's guide to agent training data draws a similar line between workflow records, action trajectories and practice environments.

Who says they buy it

We list only what buyers state on their own pages, read on November 17, 2026. A statement on a web page is not an offer. Each buyer decides case by case.

micro1 has a section titled "Monetize data from your agents". It says every agent task produces a trajectory that is valuable training data, and that it licenses qualifying trajectories for $1M to $2M annually, "with your approval". The same section describes optional agent build-outs and ongoing agent evaluations. A micro1 blog post dated October 9, 2026 says it uses de-identified company data to build reinforcement learning environments where models and agents learn to complete tasks.

Troveo lists four products it builds from a company's data. The fourth is "AI performance data", described as "AI at work in your environment". Its page does not give a price for that product.

DataPaid lists "AI feedback" among the things AI companies buy and describes it as a company's team rating AI answers on real business tasks. That is close to an evaluation transcript.

The other buyer pages we read describe operational records such as tickets, chat and documents, and do not mention agent logs. That does not mean they would decline them. Ask. Our comparison of companies that buy data for AI lists each buyer's published terms, and the micro1 comparison page sets its terms next to ours.

What is inside a trajectory

A trajectory holds more than its owner tends to expect. The agent was given access to real systems so it could do real work, and the log kept what passed through. The table lists the usual parts and the main concern each one raises for a sale.

Two parts need particular care. Tool results often contain whole customer records, because the agent fetched them to answer a question. Credentials appear because agents are handed keys and tokens to call tools, and frameworks often log the full request.

PartWhat it holdsMain concern before licensing
System promptThe standing instructions, policies and business rules given to the agentTrade secrets and internal policy wording
User inputThe request from a customer or employee, often in free textPersonal data, and anything the person typed
Model outputThe agent's messages, plans and intermediate reasoningMay repeat personal or confidential details from other parts
Tool callsThe function name and arguments sent to each systemInternal system names, account identifiers, query text
Tool resultsWhat each system returned: records, files, search results, page contentCustomer data, third-party content, client-owned records
CredentialsAPI keys, tokens, session cookies and connection strings in headers or argumentsMust be removed and rotated. Never licensed
OutcomeWhether the task succeeded, any error, and any human correctionLow risk and high value. Keep it
Evaluation dataScores, rubrics and reviewer commentsReviewer identities and internal quality standards

Rights questions

The first question in any data sale is whether the data is yours to license. With trajectories, more parties have a possible claim than with ordinary business records. The points below are general. Contracts differ, and your counsel should read yours.

Check the model provider's terms. Your agreement with the company that supplies the model governs how you may use its outputs and its interface. Ask counsel whether those terms permit licensing logged outputs to a third party for model training.

Check the agent platform and observability vendor. If traces are captured by a third-party tool, its terms govern export. Some vendors also add their own analysis fields to a trace, and those fields may be the vendor's.

Check customer contracts. If the agent served your customers, their inputs and their records are in the log. Customer agreements often limit your use of customer data to providing the service. If you run agents for a client inside the client's systems, the traces usually belong to the client. Our guide on client data vs company data explains the split.

Check third-party content. An agent that browsed the web or read licensed databases has copied that content into its tool results. You may hold no right to pass it on.

Check employee terms. Where employees prompted the agent, their inputs are work product in most cases, but notice and policy still matter. See the guide on employee data and AI licensing.

Check earlier promises. If you told customers that their conversations with your assistant would not be used for training, that statement applies to the logs.

Privacy questions

Trajectories usually contain personal data. Users type names, addresses and account details into prompts. Tool results return records about identifiable people. Privacy law applies to that content whether a person or a model wrote the record.

Three questions cover most of the ground. Find out what your privacy notice told people about the use of their data. Find out whether the proposed use is compatible with that notice or needs fresh notice or consent. Find out whether the personal data can be removed while the trajectory stays useful.

In the United States, state laws such as the California Consumer Privacy Act treat a transfer of personal information for value as a sale, with notice and opt-out duties. In Europe, the GDPR requires a lawful basis and limits use to compatible purposes. Health, financial and children's data carry extra rules. Our guides to CCPA and GDPR give general background. They are not a substitute for advice.

De-identification in trajectories is harder than in a table. The same person can appear in a prompt, a tool argument, a tool result and the model's reply, in different formats. If one mention is replaced and another is not, the record is still identifying. micro1's research post of September 24, 2026 makes the same point about enterprise data in general: the same entity should map to the same synthetic identity across a corpus.

How to prepare agent logs for sale

Preparation follows the same order as any dataset, with two steps that matter more here: removing secrets and keeping outcomes. Work on a copy. Never edit production logs.

Several steps can be done with free tools that run in your browser. The secrets scanner finds keys and tokens. The PII scrubber replaces common identifiers. The JSONL validator checks the output format. The delivery manifest tool records a checksum for each file.

  • Inventory the agents. For each one, record what it does, which tools it can call, whose data it touches and where its logs are stored.
  • Sort by owner. Separate agents that work on your own operations from agents that serve customers or run in client systems. Start with the first group.
  • Confirm retention. Many teams keep traces for a short period only. Extend retention for the agents in scope before the logs expire.
  • Export complete runs. Keep each trajectory whole, from first instruction to final outcome, with steps in order and timestamps intact.
  • Remove secrets, then rotate them. Strip keys, tokens, cookies and connection strings. Treat any credential found in a log as exposed and replace it.
  • De-identify consistently. Replace each person and each customer with the same placeholder everywhere it appears in a run.
  • Remove third-party and client content that you have no right to license, or drop the affected runs.
  • Keep the outcome and any score. A trajectory with a known result is worth more than one without.
  • Deliver in a plain structured format, one run per line. Our guide to JSONL for AI training data covers the layout.
  • Write a data card that states the agents covered, the tools they used, the date range, the cleaning method and the known gaps.

Commercial terms to settle

Agent data is produced continuously, so buyers may propose a recurring arrangement instead of a single purchase. micro1's page describes annual licensing with the company's approval. A recurring deal needs its own answers on four points.

Approval. Decide whether you approve each batch before it leaves, or approve a category once. Per-batch approval is slower and safer. Scope. A trajectory shows how your business handles its work. Ask what uses are permitted and whether the buyer may use the data to build something that competes with you. Exclusivity. Decide whether you may license the same traces to others. Our guide on exclusive vs non-exclusive licenses explains the trade.

Payment. A recurring fee needs a trigger and a clock for each period. Our guide on how AI data buyers pay sellers covers the options, and the AI data licensing agreement guide covers the contract.

Some offers bundle data licensing with services, such as building or evaluating your agents. Bundles can be useful. Price each part separately so you can tell what you are paid for the data and what you receive in services.

Outrich Data buys agent trajectories directly from the company that produced them, after a data title search. Its payment is funded into escrow before delivery. The AI agent trajectories data type page lists what qualifies, and the data valuation calculator gives an illustrative range.

Common questions

What is an AI agent trajectory?
It is the recorded sequence of steps an AI agent took to complete a task: the instruction, each model message, each tool call and result, and the final outcome. It is also called a trace or an agent log.
Who buys AI agent trajectories?
On November 17, 2026, micro1's page stated that it licenses qualifying trajectories for $1M to $2M annually with the company's approval, and Troveo's page listed AI performance data as a product. Outrich Data also buys them directly. Other buyers may consider them on request.
Can we sell logs from an agent that talks to our customers?
Possibly, after review. Those logs contain customer inputs and records, so your customer contracts, privacy notice and any promise about training use all apply. Many companies start with agents that work only on internal operations.
Do agent logs contain credentials?
Often, yes. Frameworks may log full tool requests, including API keys, tokens and connection strings. Scan for them, remove them and rotate every credential found. Credentials are never part of a licensed dataset.
Do we own the outputs our agent generated?
That depends on your agreement with the model provider and any platform vendor, and on what the agent copied into its results. Ask counsel to read those terms before you offer the logs for licensing.

Find out if your data qualifies.

Eleven questions, about five minutes, no call required. You get an answer on screen and a written follow-up.

esc