# Agent Traces: Licensing Production AI Agent Runs (2026)

> Companies running AI agents in production can license complete agent traces (task, steps, tool calls, outcome). What buyers want, legal limits and steps.

Canonical: https://origindatapartners.com/guides/license-ai-agent-traces
By: Origin Data Partners
Status: Published
Published: 2026-11-14
Updated: 2026-11-14
Sources checked: 2026-11-14

## Quick answer

Agent traces are complete records of an AI agent's runs: the task, every step and tool call with its result, the reasoning and the outcome. A company can license them if the logs are its own. Buyers want long, real production runs with a checkable outcome, the model and harness stated, and nothing cut from the middle. Runs on open-weight models are easiest, because the major providers' terms restrict using outputs to build competing models. No closed price is public.

## Key points

- An agent trace is the whole run, not the chat: task, steps, tool calls and results, reasoning and a verified outcome.
- Long non-coding runs on the company's own systems are the scarcest; generic coding traces are increasingly free.
- OpenAI, Anthropic and Google restrict using their outputs to develop competing models, so counsel reads those terms first.
- Customer content inside tool results usually belongs to, or is restricted by, the customer's contract.
- Start with 5 to 10 sample runs, or just the model and harness you use. No files are needed to begin.

## What is an agent trace, and how is it different from a chat transcript?

An [agent trace](https://origindatapartners.com/glossary/agent-trace) records one complete run of an AI agent from the task it was given to the result it produced. Researchers also call a run a [trajectory](https://origindatapartners.com/glossary/agent-trajectory). A chat transcript keeps the messages a person saw. A trace also keeps what happened between them: each [tool call](https://origindatapartners.com/glossary/tool-call) with its arguments and result, the model's reasoning where it was captured, retries and dead ends, and how the run ended.

Open formats now describe the same parts. Harbor's Agent Trajectory Interchange Format "records an agent's complete interaction history as JSON: messages, reasoning, tool calls, observations, and metrics" ([ATIF](https://docs.harborframework.com/agents/atif)), and Letta's trajectory package "formats trajectories across harnesses in a consistent, agent-friendly format" ([Letta](https://www.letta.com/blog/trajectory/)).

| Part of a run | What it holds | In a chat transcript? |
| --- | --- | --- |
| Task | The instruction or ticket the agent started from | Usually |
| Steps | Every model call, in order, with timestamps and token counts | No |
| Tool calls and results | Tool name, arguments and the full result returned (log lines, records, files) | No |
| Reasoning | The model's thinking between steps, in full, summarised or absent | No |
| Human interventions | Approvals, corrections and hand-offs | Sometimes |
| Outcome | Success, partial or failure, and who or what confirmed it | Rarely |
| Model and harness | Model name and version, the agent framework and its version, the tools available | No |


## What do buyers look for in agent traces?

AI developers and evaluation companies describe the same few requirements. A set that meets them is far easier to license than a larger set that does not.

- Real production runs on the company's own systems and data. Synthetic or scripted traffic is excluded.
- Long, multi-step runs, complete from task to outcome. The middle steps are where agents re-read, choose a wrong tool and recover, so a run with the middle cut out cannot teach recovery or build an evaluation.
- Reasoning included and nothing truncated: full tool results, not references or the first few hundred characters.
- An outcome that can be checked: a closed ticket, a reconciled ledger entry, a passing test, a merged change or a human sign-off, with its source recorded.
- Model, version and [harness](https://origindatapartners.com/glossary/agent-harness) stated, with the tools that were available to the agent.
- Runs on [open-weight models](https://origindatapartners.com/glossary/open-weight-model) where possible. Their licences can allow training on outputs; DeepSeek-R1's card says the series supports "distillation for training other LLMs" ([DeepSeek-R1](https://huggingface.co/deepseek-ai/DeepSeek-R1)).
- Logs the company owns and can license, with customer contracts, privacy notices and model terms checked first.

Why reasoning matters: Anthropic reports that "A model trained on transcripts with summarized reasoning scored 71.8% on SWE-bench Verified, compared to 80.4% for a model trained on the full reasoning" ([Anthropic](https://www.anthropic.com/news/distillation-defenses)). Proprietary APIs rarely return that full reasoning. OpenAI's guide says "reasoning tokens are not visible via the API" ([OpenAI reasoning guide](https://developers.openai.com/api/docs/guides/reasoning)); Claude's docs say "the text in a thinking block is a summary of Claude's reasoning" ([Claude thinking docs](https://platform.claude.com/docs/en/build-with-claude/thinking)), and on Claude 5 and later models the thinking display defaults to omitted until you opt in; Gemini offers "Thought summaries", but "By default, only the final output is returned" ([Gemini thinking docs](https://ai.google.dev/gemini-api/docs/thinking)). A data card should say whether each run's [reasoning tokens](https://origindatapartners.com/glossary/reasoning-tokens) are full, summarised or absent.


## Who is asking for agent traces, and what has been offered?

Each row is public evidence, graded by its source and read on 14 November 2026. A figure on a buyer's own page is that company's offer or range, a ceiling rather than a price paid. No closed price for production agent traces has been published.

| Evidence | What it says | Grade |
| --- | --- | --- |
| micro1, data partnerships page | "Every agent task produces a trajectory which is valuable training data for showing how real work gets done. We license qualifying trajectories for $1–2M annually, with your approval." ([micro1](https://company.micro1.ai/register/?referralCode=24823479-4339-4311-a157-2600b8cf72cf&utm_source=referral&utm_medium=share&utm_campaign=job_referral)) | Company claim (its own page); an offer, up to its stated figure |
| Mercor listing, as summarised by DataPacts | The listing "asks for complete session logs from a product in production for at least six months", with "a deal range of USD 20,000 to 480,000"; it excludes "Synthetic or scripted traffic" and "Truncated or partial traces" ([DataPacts](https://datapacts.com/data/trajectories/)) | Secondary source summarising a company listing |
| TraceDance paper | Built from real deployment traces: "252,557 sessions" produced "107 benchmarks", and "Nine frontier LLMs achieve a mean pass rate of only 26.7%" ([arXiv 2609.33295](https://arxiv.org/abs/2609.33295)) | Paper (preprint) |
| Exgentic release | "We study models on mountains of public data, but agents on a handful of narrow, fragmented records." ([Exgentic](https://huggingface.co/blog/michalshmu/exgentic-agent-llm-traces-v2)) | Practitioner article |
| Measuring Agents in Production (counter-signal) | Of "86 deployed systems", "68% execute at most 10 steps before human intervention" ([arXiv 2512.04123](https://arxiv.org/abs/2512.04123)) | Paper (preprint) |

Read together: real production traces make harder evaluations than public sets, buyers say they want them, and long autonomous runs are still a minority, which keeps complete long runs scarce. Published figures are ceilings set by each buyer for what it is looking for; they are not offers to any particular company.


## Which teams hold the most valuable agent traces?

Value follows three things: how long and complete the runs are, how clearly the outcome is recorded, and how much of the content the company can license. Coding traces are the most common and face the most free substitutes; long non-coding runs on real systems are the scarcest.

| Function | Typical run | Checkable outcome | What sits inside the trace |
| --- | --- | --- | --- |
| Incident investigation (SRE and on-call agents) | Reads logs, metrics and deploy history; forms a root cause | Postmortem, ticket closed, fix merged | Hostnames, secrets in environment dumps, sometimes customer IDs |
| Invoice reconciliation and finance operations | Pairs invoices with purchase orders, chases exceptions | Entry reconciled, approved or posted | Vendor bank details, invoices, staff names |
| Support queues | Triage, refund, account change; long only on escalations | Ticket closed, no reopen | Customer personal data and order history |
| Back-office paperwork | Reads documents, fills forms, files them | Filed, accepted or rejected | Personal data on forms |
| Scheduling | Checks calendars and constraints, books | Booked and confirmed | Names, contact details, locations |
| Coding | Issue to change | Tests pass, change merged, reviewer approval | Source code, keys in configuration |

In Origin's view the strongest sets for the next one to two years are non-coding runs with an outcome label from the company's own systems, ideally on open-weight models with full reasoning kept. Traces from older, weaker models lose training value quickly; their value for evaluation lasts longer when the failures they show are still hard for current models.


## Who can license agent traces, and who needs consent?

Holding traces in a tracing tool is not the same as owning everything inside them. The table is the usual position; your contracts decide the real one.

| Company type | Who controls the content | Can it license cleanly? |
| --- | --- | --- |
| Enterprise running internal agents on its own systems (SRE bot, finance operations, IT help desk) | The enterprise | Usually the cleanest, once employee notice, secrets and model terms are checked |
| AI-native company with agents on its own operations | The company | Often clean and quick; volumes are smaller |
| Agent vendor whose product runs for customers (support, SRE, accounts payable) | Each customer, for the customer content in tool results; the vendor, for its prompts, tool schemas and orchestration | Only runs on the vendor's own data, runs for customers who opt in in writing, or traces with customer content removed |
| Outsourcer or managed-service provider running agents for clients | The client, under most service contracts | Rarely without each client's written consent |
| Observability or evaluation platform holding customers' traces | The platform's customers | No: the platform processes the data for its customers |
| Any holder of regulated data (health, card, financial, privileged) | Depends on the rule | Only with the special handling in the legal section below |


## Can you license traces from Claude, GPT or Gemini? What the provider terms say

This page is general information, not legal advice. It quotes each clause and rule with its source and the date it was read; your counsel decides how they apply. A lawyer should read your own model-provider agreement (including any enterprise terms), your customer contracts and your privacy notices before any trace leaves the company.

All three providers leave outputs with the customer, and all three restrict using the services or outputs to develop competing models. Each clause below was read on its live page on 14 November 2026. What the clause means for your traces, and whether it reaches the company you would license them to, is a question for a lawyer reading your own agreement, including any enterprise terms that replace these.

| Provider and document | Clause (exact words) | Section, effective date, date read |
| --- | --- | --- |
| OpenAI Services Agreement (APIs and business services) | Customer will not, and will not permit End Users to: "(e) except for a Permitted Exception, use Output to develop artificial intelligence models that compete with OpenAI's products and services" ([OpenAI Services Agreement](https://openai.com/policies/services-agreement/)) | Section 3.3(e), Restrictions; effective January 1, 2026; read 14 November 2026 |
| Anthropic Commercial Terms of Service | "Customer may not and must not attempt to (a) access the Services to build a competing product or service, including to train competing AI models or resell the Services except as expressly approved by Anthropic; (b) reverse engineer or duplicate the Services; or (c) support any third party's attempt at any of the conduct restricted in this sentence." ([Anthropic Commercial Terms](https://www.anthropic.com/legal/commercial-terms)) | Section D.4, Use Restrictions; effective June 17, 2025; read 14 November 2026 |
| Gemini API Additional Terms of Service | "You may not use the Services to develop models that compete with the Services (e.g., Gemini API or Google AI Studio)." ([Gemini API terms](https://ai.google.dev/gemini-api/terms)) | Use Restrictions; effective March 23, 2026; read 14 November 2026 |

What the clauses say, without interpreting them: OpenAI defines a "Permitted Exception" narrowly (internal models that categorize, classify or organize data and are not distributed, and fine-tuning through OpenAI's own services). The OpenAI agreement says the customer "owns all Output"; Anthropic's says the customer "owns its Outputs", and its assignment applies "Subject to Customer's compliance with these Terms"; Google says it "won't claim ownership over that content". Ownership of outputs and permission to use them for a competing model are separate questions in each document. Anthropic's clause also covers supporting "any third party's attempt" at the restricted conduct (limb (c)); counsel should read it alongside limb (a) when traces would go to another company.

The consumer OpenAI terms are shorter: users may not "Use Output to develop models that compete with OpenAI" ([OpenAI Terms of Use](https://openai.com/policies/terms-of-use/)). Using traces only to evaluate a model is a different use from training one, but none of the three documents carves evaluation out, so do not assume it is safe. Runs on open-weight models avoid these clauses, though each model's own licence still applies: Qwen3-8B is "apache-2.0" ([Qwen3-8B](https://huggingface.co/Qwen/Qwen3-8B)) and gpt-oss-120b has a "Permissive Apache 2.0 license" ([gpt-oss-120b](https://huggingface.co/openai/gpt-oss-120b)). Check the licence of the exact model version you ran.


## What legal limits apply to the data inside agent traces?

Every tool call that reads a CRM, an inbox, a ticket or a database copies that content into the trace. Hugging Face's own guidance for traces warns that they "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)). The limits below decide what can leave the company and in what form; the [agent trace redaction checklist](https://origindatapartners.com/guides/agent-trace-redaction-checklist) turns them into steps.

- Customer data and vendor contracts: customer content in tool results is usually governed by the customer's master services agreement and data processing agreement, which often limit processing to providing the service. Exclude it, get written opt-in, or remove it before licensing. Third-party data inside tool results (enrichment, market data) is licensed to you, not resellable.
- Employee data: traces record who approved, who was paged and staff messages. Employees need notice under privacy law, and works councils may need to be consulted in some EU countries.
- Regulated data: health information needs HIPAA de-identification; Safe Harbor removes the identifiers listed in 45 CFR 164.514(b)(2), from "(A) Names" onward ([45 CFR 164.514](https://www.law.cornell.edu/cfr/text/45/164.514)). Card verification codes are sensitive authentication data that "cannot be stored after authorization" ([PCI SSC](https://www.pcisecuritystandards.org/faq/articles/Frequently_Asked_Question/can-card-verification-codes-values-be-stored-for-card-on-file-or-recurring-transactions/)), so a trace that holds one is a problem before it is a licensing question. Privileged legal material stays out.
- Privacy law: under GDPR, data that has been pseudonymised "should be considered to be information on an identifiable natural person" ([GDPR Recital 26](https://gdpr-info.eu/recitals/no-26/)), so consistent pseudonyms reduce risk but do not end the analysis. US state privacy laws apply to the personal data of their residents.
- Confidentiality and trade secrets: system prompts, tool schemas and internal runbooks can be trade secrets, and NDAs or client work-product clauses can bar disclosure. Decide what to keep, describe or remove before a sample is prepared.


## What goes on an agent trace data card?

A data card describes the set one run at a time, so a buyer can judge it without opening a single trace. [Download the data card template (CSV)](https://origindatapartners.com/downloads/agent-trace-data-card.csv); it has one row per field and two fictional example runs. Offer the runs as JSONL (one run per line), in an open trajectory format, or as OpenTelemetry GenAI spans with a run index.

| Field | What to record |
| --- | --- |
| run_id, started_at, ended_at | A stable ID and timestamps with time zone |
| task_text | The instruction as given, redacted |
| function | For example incident investigation or invoice reconciliation |
| model, model_version, provider, open_weight | Exactly what ran, and whether the weights are open |
| harness, harness_version, tools_available | The agent framework and the tool names and schemas the agent could call |
| steps | Each step's type (reasoning, tool call, tool result, message), content, tool name, arguments, result, latency and tokens |
| reasoning_visibility | Full, summary or none |
| token_counts | Input, output, reasoning and total |
| outcome, outcome_source | Success, partial or failure, and what confirmed it (human sign-off, ticket state, test result, ledger) |
| human_interventions | Corrections and approvals, with the step they came at |
| redaction_log | Entity types removed, method, counts and reviewer |
| licence_scope, consent_basis | What the run may be used for, and why the company can license it |


## Agent traces checklist: what makes a set licensable?

Write these down before a first conversation. None of them needs a file to leave the company.

| Check | What to write down |
| --- | --- |
| Real runs | Production runs on your own systems; no synthetic or scripted traffic |
| Length and completeness | Typical steps and tokens per run; whether tool results or middle steps were truncated or sampled |
| Reasoning | Full (open-weight models), summarised or absent, per model |
| Outcome label | How each run was verified, and where that record lives |
| Model and harness | Model names and versions; framework and version; tracing tool and retention |
| Ownership | Whose content appears in tool results; customer contracts, DPAs, privacy notices and model terms checked |
| Redaction | Secrets found and rotated; personal and customer data removed and logged |


## What does a first sample of agent traces look like?

Five to ten complete runs, redacted, with the model and harness stated and the outcome of each run recorded. That is enough for a buyer to judge length, completeness and domain. Nothing should leave the company before the rights and redaction checks below; until then, a written description of the runs works just as well.

In Origin's view, a few hundred complete, long, outcome-labelled runs is a realistic first set. Because each new model and harness produces new runs, a set can also be refreshed, which suits recurring licences better than a one-off export.


## What happens next: three steps

A first conversation needs no files. You can stop after any step.

- Describe your runs: tell Origin your model and harness and describe 5 to 10 sample runs, with no files and no system access ([describe your agent traces](https://origindatapartners.com/check-fit?sector=agent-traces#company-review)).
- Rights and redaction check: Origin works through the provider terms, customer contracts and privacy limits with you and your counsel, then the redaction plan and the data card. You approve a small redacted sample before anything leaves the company.
- Buyer review and terms: Origin works with buyers, the AI developers and data buyers that license records like these, and shares the data card only with buyers you approve. You approve scope, use, price and delivery, and sign the licence yourself.
- Origin does not charge data owners; a buyer may pay Origin when a deal closes ([how Origin is paid](https://origindatapartners.com/referral-disclosure)).


## Where do your agent traces live, and how do you export them?

Most teams already keep traces for debugging. The [guide to exporting agent traces](https://origindatapartners.com/guides/export-agent-traces) gives the documented export steps, limits and retention windows for each tool below.

| Where traces sit | What to check first |
| --- | --- |
| LangSmith, Langfuse, Arize Phoenix, Braintrust, Helicone, W&B Weave | Plan retention and whether bulk export is on your plan |
| Datadog LLM Observability | Retention by plan; the export API defaults to the last 15 minutes |
| OpenTelemetry GenAI spans in any backend | Whether content capture was switched on when the agent ran |
| OpenHands conversations, LangGraph checkpoints | Event streams and state history, joined to model-call spans |


## Guides for the team that keeps the traces

Practical pages for platform and ML engineers, and the related guides behind this one:

- [How to export agent traces from each tracing tool](https://origindatapartners.com/guides/export-agent-traces)
- [Agent trace redaction checklist](https://origindatapartners.com/guides/agent-trace-redaction-checklist)
- [AI usage data and agent logs from AI products](https://origindatapartners.com/guides/license-ai-product-usage-data)
- [Data for AI evaluation: training, testing and holdouts](https://origindatapartners.com/guides/license-data-for-ai-evaluation)
- [RL environments built from company data](https://origindatapartners.com/guides/rl-environments-company-data)
- [Workflow data for AI agents](https://origindatapartners.com/guides/license-workflow-data-for-ai-agents)

Running agents in production? [Describe 5 to 10 sample runs, or just tell us your model and harness](https://origindatapartners.com/check-fit?sector=agent-traces#company-review). No files or system access needed.

Origin Data Partners works with companies that own data and helps get it bought. We are independent of the services on this page, and a buyer may pay us when a deal closes ([how Origin is paid](https://origindatapartners.com/referral-disclosure)). [See what your data is worth](https://origindatapartners.com/check-fit).


## Questions and answers

### Can we license agent traces from runs on Claude, GPT or Gemini?

Only after counsel reads your agreement. OpenAI's Services Agreement (3.3(e)), Anthropic's Commercial Terms (D.4) and the Gemini API terms each restrict using the services or outputs to develop competing models, while leaving the outputs with you. Runs on open-weight models avoid those clauses, subject to the model's own licence.

### Do we need customer consent to license agent traces?

When tool results carry customer content, usually yes, or that content has to come out first. Customer contracts and data processing agreements often limit processing to providing the service. Runs over your own internal data need no customer consent, though employee notice and model terms still apply.

### How many agent runs is enough to start?

Five to ten complete sample runs, with the model and harness stated, are enough for a first look. In Origin's view a few hundred long, complete, outcome-labelled runs make a realistic first set, and sets that refresh as models change suit recurring licences.

### Does it matter if the model's reasoning was hidden or summarised?

Yes. Full reasoning is worth more for training. OpenAI says reasoning tokens are not visible via its API; Claude returns at most a summary of its thinking, and on Claude 5 and later models nothing unless you opt in; Gemini returns thought summaries only when you enable them. Record full, summary or none for every run instead of leaving it out.

### Are failed agent runs worth anything?

Often yes, especially for evaluation. A failure with its full middle steps shows where agents go wrong, as long as it is labelled as a failure and the outcome source is recorded.

### Will licensing traces expose our prompts and tools?

A trace includes system prompts and tool schemas unless you remove them. Confidential instructions can be redacted or replaced with a description, and the licence can limit use and onward sharing. Decide this before any sample is prepared.

### Can we keep training our own model on the same traces?

Under a non-exclusive licence, yes. An exclusive licence can restrict your own use or other licences, so read its scope. Many teams already fine-tune on their traces in-house; that is a separate use of the same data and compatible with a non-exclusive licence.

### How much are production agent traces worth?

No closed price for production agent traces is public. micro1's page says it licenses qualifying trajectories "for $1–2M annually, with your approval" (read 14 November 2026), and a secondary summary of a Mercor listing gives USD 20,000 to 480,000 per deal. Both are ranges set by the buyer, ceilings rather than prices paid.

### What does Origin charge to review agent traces?

Nothing. Origin does not charge data owners; a buyer may pay Origin when a deal closes. You approve every step and sign the licence yourself.

## Sources

- [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)
- [OpenAI Terms of Use (consumer, effective January 1, 2026)](https://openai.com/policies/terms-of-use/)
- [OpenAI API docs: reasoning models](https://developers.openai.com/api/docs/guides/reasoning)
- [Claude API docs: thinking](https://platform.claude.com/docs/en/build-with-claude/thinking)
- [Gemini API docs: thinking](https://ai.google.dev/gemini-api/docs/thinking)
- [micro1: data partnerships page (read 14 November 2026)](https://company.micro1.ai/register/?referralCode=24823479-4339-4311-a157-2600b8cf72cf&utm_source=referral&utm_medium=share&utm_campaign=job_referral)
- [DataPacts: trajectories page summarising a Mercor listing (secondary source)](https://datapacts.com/data/trajectories/)
- [TraceDance: benchmarks from real-world agent deployment traces (arXiv 2609.33295)](https://arxiv.org/abs/2609.33295)
- [Measuring Agents in Production (arXiv 2512.04123)](https://arxiv.org/abs/2512.04123)
- [Exgentic agent traces release (Hugging Face community article, 4 October 2026)](https://huggingface.co/blog/michalshmu/exgentic-agent-llm-traces-v2)
- [Anthropic: distillation defenses (12 October 2026)](https://www.anthropic.com/news/distillation-defenses)
- [Harbor: Agent Trajectory Interchange Format (ATIF)](https://docs.harborframework.com/agents/atif)
- [Letta: the trajectory package (23 July 2026)](https://www.letta.com/blog/trajectory/)
- [Hugging Face Hub docs: agent traces](https://huggingface.co/docs/hub/en/agent-traces)
- [DeepSeek-R1 model card (MIT License)](https://huggingface.co/deepseek-ai/DeepSeek-R1)
- [Qwen3-8B model card (Apache 2.0)](https://huggingface.co/Qwen/Qwen3-8B)
- [gpt-oss-120b model card (Apache 2.0)](https://huggingface.co/openai/gpt-oss-120b)
- [45 CFR 164.514(b)(2): HIPAA Safe Harbor identifiers](https://www.law.cornell.edu/cfr/text/45/164.514)
- [PCI Security Standards Council FAQ: card verification codes and sensitive authentication data](https://www.pcisecuritystandards.org/faq/articles/Frequently_Asked_Question/can-card-verification-codes-values-be-stored-for-card-on-file-or-recurring-transactions/)
- [GDPR Recital 26: pseudonymised data stays personal data](https://gdpr-info.eu/recitals/no-26/)

## Related reading

- [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)
- [Can an AI startup license its usage data and agent logs?](https://origindatapartners.com/guides/license-ai-product-usage-data)
- [License data for AI evaluation: training, testing and holdouts](https://origindatapartners.com/guides/license-data-for-ai-evaluation)
- [What are RL environments? How AI labs build them and why they pay for company data](https://origindatapartners.com/guides/rl-environments-company-data)
- [License workflow data for AI agents: a company owner’s guide](https://origindatapartners.com/guides/license-workflow-data-for-ai-agents)
- [AI Training Data Price Index: Disclosed Deals (2026)](https://origindatapartners.com/guides/ai-training-data-price-index)
- [/glossary/agent-trace](https://origindatapartners.com/glossary/agent-trace)
- [/glossary/open-weight-model](https://origindatapartners.com/glossary/open-weight-model)