Skip to content
Outrich Data

Data type

License your AI agent logs and trajectories for AI training

Logs from AI agents your company runs can often be licensed for AI training once credentials, customer data and third-party content are removed. Outrich Data is the direct buyer. It checks model provider, platform and customer terms, strips secrets and identifiers, and pays you for a licensed copy that you approve in writing before delivery.

What counts as ai agent trajectories

An agent trajectory is the full record of one AI agent run. It holds the instruction the agent was given, each message the model produced, every tool call with its arguments, what each tool returned, and how the task ended. Agent frameworks and observability tools write these records automatically. The same family includes tool-call traces, which are the structured requests and responses, and evaluation transcripts, which are runs that a person or an automated check has scored.

The useful archive covers agents doing real work on your own operations: triaging tickets, reconciling records, writing and testing code, drafting replies for staff to approve. Runs with a known result and any human correction are the core of it. Traces from agents you operate for a client inside the client's systems belong to that client and are excluded. So are test runs against synthetic prompts, which show little about real work.

Why AI labs want it

AI labs are training models to complete multi-step tasks in software, and a trajectory is a direct example of one. It shows which tool was chosen, in what order, with what input, and what came back. Public data has almost none of this, because agent runs happen inside company systems and most teams delete the traces after a short retention period.

Failures and corrections matter as much as successes. A run where the agent called the wrong tool, looped, or was overruled by a reviewer shows a model what to avoid. Scored runs let a lab build evaluations in which success is checked against what really happened. Trajectories tied to the business outcome of the task are the most useful of all.

Removed before licensing

What comes out first.

  • API keys, tokens, session cookies and connection strings, which are also rotated
  • Customer names, contact details and account identifiers in prompts and tool results
  • Whole customer records an agent fetched to answer a question
  • Employee and reviewer names, replaced with role labels
  • Third-party content the agent copied from websites or licensed databases
  • Internal hostnames, file paths and system names that expose your infrastructure
  • System prompt passages that state confidential policy or pricing rules, where you choose to withhold them

Data title search

What gets checked.

  • Model provider terms. Your agreement with the model supplier governs use of its outputs. Counsel should confirm it permits licensing logged outputs for training.
  • Platform and observability vendor terms. Check the right to export traces, and whether any analysis fields in a trace were generated by the vendor.
  • Customer contracts. Where an agent served customers, their inputs and records are in the log. Many agreements limit use of customer data to providing the service.
  • Privacy notices and earlier statements. If you told users their conversations with an assistant would not be used for training, that applies to the logs.
  • Third-party sources. Content an agent retrieved from the web or a licensed database may not be yours to pass on.
  • Employee notice. Staff who prompted or reviewed the agent should have been told how those records may be used.

Who signsThe head of engineering or AI sponsors it, with the security lead confirming that secrets were removed and rotated. The general counsel approves after reviewing model provider and customer terms.

Value

What moves the price of ai agent trajectories.

FactorEffect on value
Known outcomesRuns that record whether the task succeeded, and any human correction, are generally worth more than runs that end without a result.
Real tasksAgents working on live operations are scarcer than demos and test runs against invented prompts.
Tool depthRuns that span several tools and many steps show more of how work is done than single-call exchanges.
Complete tracesEvery step in order, with arguments and results intact after cleaning, is more useful than summaries or final answers alone.
Clean rightsAgents that touch only your own data and systems are simpler to clear than agents that handle customer or client records.

Industries where this is common

Questions

Licensing ai agent trajectories: what companies ask.

Can we license logs from an agent built on a third-party model?
That depends on your agreement with the model provider and any platform in between. Outrich Data asks for those terms in the data title search, and your counsel decides whether they permit licensing the logged outputs.
Our agent logs contain API keys. Is that a problem?
It is common and it is fixable. Keys, tokens and connection strings are removed before any package is built, and you should rotate every credential found. The free secrets scanner runs in your browser and can show you the scale of the issue first.
How are agent trajectories different from workflow logs?
Workflow logs record what people and software did in a system of record. A trajectory records what a model did, including its own messages and tool calls. Many companies hold both, and the two can be licensed together.
Is this a one-time sale or an ongoing arrangement?
Either. Agents produce new runs every day, so a license can cover a historical archive, later refreshes or both. Each delivery is a package you approve in writing. The guide to selling AI agent logs and trajectories covers the terms to settle.

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