# Non-Coding Agent Traces: What Buyers Ask For (2026)

> Support, invoice, incident and scheduling agents leave full run logs. What buyers ask for, what to send first, how to export and what to remove.

Canonical: https://origindatapartners.com/guides/license-non-coding-agent-traces
By: Origin Data Partners
Status: Published
Published: 2026-11-15
Updated: 2026-11-15
Sources checked: 2026-11-15

## Quick answer

Non-coding agent traces are complete runs of AI agents that do business work: a support queue, invoice reconciliation, incident investigation, scheduling or back-office paperwork. Buyers ask for real production runs with every step, tool call, result and the model's reasoning, each long and finished with a way to tell it worked. Coding runs count too, but non-coding runs are the bigger gap. Start with 5 to 10 sample runs plus the model and harness.

## Key points

- A trace is the whole run: task, every step, tool calls and results, reasoning and the result. A chat transcript is not enough.
- Buyers ask for real production runs on your own systems, about 100K tokens or more each, with nothing cut from the middle.
- Each run needs a way to tell it finished the job: a human sign-off, a closed ticket or a checked result.
- Open-weight models are a plus; any model is accepted if the model and harness are stated.
- Remove secrets, customer content and personal data before any copy leaves the company.

## What counts as a non-coding agent trace?

One complete run of an AI agent doing business work, from the task it was given to the result it produced. It includes every step, each [tool call](https://origindatapartners.com/glossary/tool-call) with its arguments and the full result, the model's reasoning and how the run ended.

A chat transcript keeps only the messages a person saw. A support agent's transcript shows the customer's question and the reply; its trace also shows the order lookup, the refund policy it read, the refund call it made and whether the ticket stayed closed. That middle is what a model learns from. The [guide to licensing agent traces](https://origindatapartners.com/guides/license-ai-agent-traces) lists every part of a run in a table.


## What do buyers ask for in agent traces?

Full runs on real work, long, complete and with a known result. This list is what buyers Origin works with asked for in November 2026. It is Origin's own account of those requests, not a quote from a public page, and no buyer is named.

- Full runs, not chat transcripts: the task, every step, the tool calls and their results, the model's reasoning and the result.
- Real production runs on the company's own systems, not demos or scripted tests.
- Long, multi-step runs, about 100K tokens or more each.
- Reasoning included, with nothing cut from the middle of the run.
- Runs that finished the job, with a way to tell: a human sign-off, a closed ticket or a checked result.
- Open-weight models (DeepSeek, Kimi, GLM, Qwen, Llama) are a plus. Any model is accepted if the model and harness are stated.
- Logs the company owns and can license.
- A few hundred runs is enough to start. Send 5 to 10 sample runs plus the model and harness first.

Volume alone does not help. In the words of a trace buyer's published essay (1 October 2026), "Many production traces will be repetitive, incomplete, or too basic to contain content valuable enough for frontier model development." The same essay notes that "Production sessions often contain customer information and proprietary material so permissions and appropriate filtering inform which tasks we can construct."


## What does a good run look like in support, finance, incidents and the back office?

Long, tool-heavy and ending in a record you can check. The table shows the typical run in each function, the record that proves it worked and what sits in the tool results that has to be dealt with first.

| Function | Typical long run | How the outcome is proven | What sits in tool results | Licensing difficulty |
| --- | --- | --- | --- | --- |
| Support queue | Escalations: account changes, refunds, multi-system lookups | Ticket closed and not reopened; refund posted | Customer names, order history, messages | High: customer personal data throughout |
| Invoices, payments and reconciliation | Pair each invoice with its purchase order and receipt, chase exceptions, prepare a payment batch | Line reconciled, payment released after approval, no reversal | Vendor bank details, invoice amounts, staff names | Medium: bank data must come out |
| Incident investigation | Read alerts, logs, metrics and deploy history; name a root cause | Root cause accepted in the postmortem, fix merged, alert cleared | Hostnames, secrets in environment dumps, customer IDs in logs | Medium: secrets first |
| Scheduling | Check calendars, constraints and rules; book and rebook | Booking confirmed and kept | Names, contact details, locations | Medium |
| Back-office paperwork | Read documents, fill forms, file them | Filing accepted, not returned | Personal data on forms | Medium to high |
| IT service desk | Diagnose, change access or settings, close the request | Request closed, no repeat request | Employee names, device and account details | Low to medium for internal desks |

Agents like these already run at volume. A payment batch agent in finance software "ran 14 payment batches on its own, covering 2,395 individual payments worth $20.1 million, in its first five weeks in production" ([PYMNTS](https://www.pymnts.com/news/artificial-intelligence/2026/one-ai-agent-2395-payments-still-no-final-say/)).


## Why are non-coding agent traces scarcer than coding traces?

Most agent work and most public agent data is about code, so coding runs are easy to find. Non-coding runs happen inside a company's own systems, with real customers and records, and that is hard to fake.

Researchers who mapped "43 benchmarks and 72,342 tasks" against the labour market found that agent development "tends to be programming-centric" ([arXiv 2603.01203](https://arxiv.org/abs/2603.01203)). Real business work is also still hard for models. On an incident benchmark where "Each task provides a Kubernetes incident snapshot containing alerts, events, traces, metrics, logs, and application topology", every frontier model scored under half, "making ITBench-AA SRE one of the least saturated agentic benchmarks in our suite" ([Artificial Analysis](https://artificialanalysis.ai/articles/itbench-aa-launch)). Runs that show how such work gets done on real systems are what is missing.


## What should you send a buyer first?

Five to ten sample runs, plus the model and the harness that ran them. Before any file, a written description of those runs is enough to start.

Use the checklist below for each sample run. Fill it in from your tracing tool; nothing has to leave the company to answer it.

| Field | What to write |
| --- | --- |
| Run ID | The stable ID in your tracing tool |
| Function | For example support escalation, invoice exception, incident investigation |
| Token count | Input, output and reasoning tokens for the whole run |
| Steps | Number of model calls and tool calls |
| Outcome and its source | Success, partial or failure, and what confirmed it (ticket state, ledger, postmortem, sign-off) |
| Model and harness | Model name and version; agent framework and version |
| Reasoning kept | Full, summary or none |
| Redaction status | Not started, secrets removed, fully redacted |


## Example: one sample run, described

A fictional mid-size logistics company runs an accounts payable agent on an open-weight model. Its finance systems lead describes one run in ten minutes. The figures are invented to show the method.

| Field | Value |
| --- | --- |
| Run ID | ap-2026-10-14-0412 |
| Function | Invoice exception: quantity on the invoice differs from the goods receipt |
| Token count | 142,000 total, of which 38,000 reasoning |
| Steps | 61 model calls, 44 tool calls (ERP lookups, supplier email draft, receipt query) |
| Outcome and its source | Success: line reconciled after the supplier sent a credit note; approved by an AP clerk on 16 October |
| Model and harness | Qwen model served on vLLM; LangGraph, version recorded |
| Reasoning kept | Full, in a separate reasoning field |
| Redaction status | Secrets removed; supplier bank details still present |

Fictional example. Ten rows like this, one per sample run, are a complete first description.


## How do you export runs from the tools you already use?

Most teams already keep runs for debugging. The [guide to exporting agent traces](https://origindatapartners.com/guides/export-agent-traces) covers each tracing tool in detail; the short version is below.

| Where runs live | How to get a complete run out |
| --- | --- |
| LangGraph | Checkpoints hold state per step; [the docs](https://docs.langchain.com/oss/python/langgraph/checkpointers) note that "the checkpoints will be ordered chronologically with the most recent checkpoint" first. Join them to model-call spans |
| LangSmith | Bulk export: "Export LangSmith trace data to an S3-compatible bucket in Parquet format" ([LangSmith docs](https://docs.langchain.com/langsmith/data-export)) |
| Langfuse | For one run, "open the trace and click Download JSON in the header" ([Langfuse](https://langfuse.com/faq/all/export-single-trace-or-session)); the API for many |
| Arize Phoenix | "download all spans as a dataframe using the get_spans_dataframe() method" ([Phoenix docs](https://arize.com/docs/phoenix/tracing/how-to-tracing/importing-and-exporting-traces/extract-data-from-spans)) |
| OpenHands | Trajectory download endpoint "/api/file/download-trajectory/{conversation_id}" ([OpenHands docs](https://docs.openhands.dev/sdk/guides/agent-server/api-reference/files/download-trajectory)) |
| OpenTelemetry GenAI spans (any backend) | Check content capture was on: instrumentations "SHOULD NOT capture them by default, but SHOULD provide an option for users to opt in" ([OpenTelemetry](https://github.com/open-telemetry/semantic-conventions-genai/blob/main/docs/gen-ai/gen-ai-spans.md)) |
| OpenAI Agents SDK | Traces go to OpenAI's dashboard; a processor you add "will receive traces and spans as they are ready" so you keep your own copy ([Agents SDK docs](https://openai.github.io/openai-agents-python/tracing/)) |
| Salesforce Agentforce | The Session Tracing Data Model "structures every logged event in a session down to a single conversation turn" ([Salesforce](https://www.salesforce.com/blog/agentforce-observability/)). Read your Salesforce agreement first |
| Agno, AWS Bedrock AgentCore, MLflow | See the export guide and your platform documentation; check that offloaded tool results are included |


## What must be removed from agent logs before anyone outside sees them?

Secrets first, then customer and third-party content, then personal data, card and bank data, and internal hosts. Rotate every secret you find, because removing it from the copy does not undo the exposure.

The order matters. On 13 November 2026 Cursor disclosed that a storage bucket "contained limited coding agent traces, mostly from 2025, from sessions not in Privacy Mode, which we were permitted to use for model training", and that "A small number of those traces contained credentials that had not been redacted" ([Cursor](https://cursor.com/blog/security-incident-update)). An attacker who copied the bucket used those credentials to reach other systems. Traces "can include prompts, tool inputs, command output, local paths, screenshots, secrets, private code, and personal data" ([Hugging Face](https://huggingface.co/docs/hub/en/agent-traces)).

- Secrets: API keys, tokens, passwords and connection strings, in tool arguments, results and environment dumps. Scan, remove, rotate.
- Customer and third-party content in tool results: records your customers own, data licensed from vendors.
- Personal data: names, emails, phone numbers, addresses, account numbers, in task text, results and reasoning.
- Card and bank data: card numbers, verification codes, bank account and routing numbers.
- Internal hosts, IP addresses and file paths that map your network.

The [agent trace redaction checklist](https://origindatapartners.com/guides/agent-trace-redaction-checklist) turns this into steps with the tools to use.


## Who can license non-coding agent traces?

A company running its own agents on its own systems usually can, once the checks below are done. A vendor running agents for customers, or an outsourcer running them for clients, usually needs each customer's or client's consent.

The [table of who can license](https://origindatapartners.com/guides/license-ai-agent-traces#who-can-license-agent-traces) on the hub covers each case. Runs on Claude, GPT or Gemini need counsel to read the provider's terms first: OpenAI's agreement bars using "Output to develop artificial intelligence models that compete with OpenAI's products and services" except for a Permitted Exception ([OpenAI Services Agreement](https://openai.com/policies/services-agreement/)), Anthropic's bars accessing the services "to build a competing product or service, including to train competing AI models" ([Anthropic Commercial Terms](https://www.anthropic.com/legal/commercial-terms)), and Google's says "You may not use the Services to develop models that compete with the Services" ([Gemini API terms](https://ai.google.dev/gemini-api/terms)). Check your own contract, including any enterprise terms that replace these. This page is general information, not legal advice.


## Common mistakes with non-coding agent traces

Most of them come from treating logs kept for debugging as if they were already a dataset.

- Sending chat transcripts without the tool calls and results.
- Sending runs whose middle was summarised by the harness, without saying so.
- Leaving the outcome blank because "the ticket closed".
- Not stating the model and harness for each run.
- Sharing a sample before secrets are found and rotated.
- Including runs done on a client's or customer's data without their consent.
- Counting short, routine runs toward the few hundred that matter.


## Non-coding agent runs as a data asset

A few hundred long runs of agents doing real support, finance, incident or back-office work, each with its reasoning and a checked outcome, are scarce, and buyers say so. No company has published the price it was paid for its production traces, so any figure you hear is an offer, not a market price.

Origin helps you describe the runs, check the rights, plan the redaction and build the data card from the [template](https://origindatapartners.com/downloads/agent-trace-data-card.csv). You approve each run before it leaves. We help you license your data. Origin Data Partners is an independent data partner: we tell you what your records could be worth and prepare the records, the terms and the license with you. The AI company that licenses them pays your company; a buyer may pay Origin when a deal closes.

Does your company hold records like these? [See what your data is worth](https://origindatapartners.com/check-fit).


## Questions and answers

### Can a company license the logs of its customer-support AI agent?

Often yes, for runs on its own support desk, once customer personal data is removed and the provider terms are checked. The logs must include the tool calls, results and outcome, not just the conversation. Runs a vendor operates for you depend on your contract with that vendor.

### Do support-agent conversations count as agent traces?

Only with the tool calls, their results and the outcome. A transcript of what the customer and the agent said is a chat log. The trace adds the lookups, actions and checks between the messages and whether the ticket stayed closed.

### Our agent runs on Claude or GPT. Can we still license its runs?

Possibly, but counsel reads the output clauses first. OpenAI, Anthropic and Google each restrict using their outputs to build competing models. The [provider terms section of the agent traces guide](https://origindatapartners.com/guides/license-ai-agent-traces#agent-traces-legal-limits-provider-terms) quotes each clause with its date; your own contract, including enterprise terms, decides.

### How many agent runs do we need before talking to a buyer?

Five to ten sample runs, or a written description of them, plus the model and harness. A few hundred long runs is enough for a first set.

### Our agent runs are short. Are they worth anything?

Less for training. Escalations, exceptions and multi-step cases are usually the valuable minority, so count those separately from routine runs.

### Why are non-coding agent traces worth more than coding agent traces?

Because they are scarcer. Coding runs and coding benchmarks are plentiful, while runs of agents doing support, finance or incident work on real systems sit inside companies and are hard to simulate.

### Can we keep fine-tuning our own model on traces we license?

Yes under a non-exclusive licence. An exclusive licence can limit your own use, so read its scope before signing.

## Sources

- [PYMNTS: one AI agent, 2,395 payments (21 August 2026)](https://www.pymnts.com/news/artificial-intelligence/2026/one-ai-agent-2395-payments-still-no-final-say/)
- [How Well Does Agent Development Reflect Real-World Work? (arXiv 2603.01203)](https://arxiv.org/abs/2603.01203)
- [Artificial Analysis: ITBench-AA, SRE incident tasks (27 May 2026)](https://artificialanalysis.ai/articles/itbench-aa-launch)
- [Cursor: an update on our recent security incident (13 November 2026)](https://cursor.com/blog/security-incident-update)
- [Hugging Face Hub docs: agent traces](https://huggingface.co/docs/hub/en/agent-traces)
- [Salesforce: Agentforce Observability and the Session Tracing Data Model (21 July 2026)](https://www.salesforce.com/blog/agentforce-observability/)
- [OpenAI Agents SDK docs: tracing](https://openai.github.io/openai-agents-python/tracing/)
- [LangGraph docs: checkpointers](https://docs.langchain.com/oss/python/langgraph/checkpointers)
- [LangSmith docs: bulk export trace data](https://docs.langchain.com/langsmith/data-export)
- [Langfuse FAQ: export a single trace or session](https://langfuse.com/faq/all/export-single-trace-or-session)
- [Arize Phoenix docs: extract data from spans](https://arize.com/docs/phoenix/tracing/how-to-tracing/importing-and-exporting-traces/extract-data-from-spans)
- [OpenHands agent server API: download trajectory](https://docs.openhands.dev/sdk/guides/agent-server/api-reference/files/download-trajectory)
- [OpenTelemetry GenAI semantic conventions: model spans and content capture](https://github.com/open-telemetry/semantic-conventions-genai/blob/main/docs/gen-ai/gen-ai-spans.md)
- [OpenAI Services Agreement (effective January 1, 2026), section 3.3(e)](https://openai.com/policies/services-agreement/)
- [Anthropic Commercial Terms of Service (effective June 17, 2025), section D.4](https://www.anthropic.com/legal/commercial-terms)
- [Gemini API Additional Terms of Service (effective March 23, 2026), Use Restrictions](https://ai.google.dev/gemini-api/terms)

## Related reading

- [Agent Traces: Licensing Production AI Agent Runs (2026)](https://origindatapartners.com/guides/license-ai-agent-traces)
- [Export Agent Traces: LangSmith, Langfuse, Phoenix and More](https://origindatapartners.com/guides/export-agent-traces)
- [PII Redaction for Agent Traces and LLM Logs: A Checklist](https://origindatapartners.com/guides/agent-trace-redaction-checklist)
- [License workflow data for AI agents: a company owner’s guide](https://origindatapartners.com/guides/license-workflow-data-for-ai-agents)
- [Agent Run Outcome Labels: Tickets, Ledgers, Sign-Offs](https://origindatapartners.com/guides/agent-run-outcome-labels)
- [Long-Horizon Agent Traces: How to Keep the Middle](https://origindatapartners.com/guides/long-horizon-agent-traces)
- [Open-Weight Agent Traces: Qwen, DeepSeek, Kimi, GLM](https://origindatapartners.com/guides/open-weight-agent-traces)
- [/glossary/agent-harness](https://origindatapartners.com/glossary/agent-harness)
- [/refer/ai-agent-consultants-and-builders](https://origindatapartners.com/refer/ai-agent-consultants-and-builders)
- [Export Agent Runs: Agno, AgentCore, Agentforce, OpenAI SDK](https://origindatapartners.com/guides/export-agent-runs-from-agent-platforms)
- [AI SRE Agent Traces: Incident Runs as Training Data](https://origindatapartners.com/guides/ai-sre-incident-agent-traces)
- [Vendor AI Agent Logs: Who Owns Them and What to Export](https://origindatapartners.com/guides/vendor-ai-agent-logs-ownership)
