Skip to content
Synthetic educational data

AI Agent Operations Dataset

Analyze how a fictional company's AI agents perform in production. Linked tables cover agent versions, sessions, model calls with token counts and cost, tool calls with errors, and evaluation scores.

Good for: LLM cost reporting, agent reliability analysis, version comparisons, and fan-out-safe joins.

Download this dataset

Pick the format that fits your project. Downloads are free with an account, and you can create one in the same step.

While it downloads: answer 5 questions in your browser

Run SQL on these tables right here, with no setup. Each answer is checked instantly, and every question comes with hints.

Preview the data

Explore a sample before downloading. Select a table to see its columns, row count, relationships, and what each row represents.

Models

One row represents: One row per model.

Rows in this table
6
Connects to
Agent versions, Llm calls

Sample rows · download for the full table

Sample rows from models. This preview is not the full download.
model_idmodel_nameprovidertierinput_price_per_mtokoutput_price_per_mtokcontext_window
MDL01atlas-largeMeridian Labsflagship3.0015.00200000
MDL02atlas-smallMeridian Labsfast0.251.25200000
MDL03nimbus-proCirrus AIflagship2.5010.00128000
MDL04nimbus-miniCirrus AIfast0.150.60128000
MDL05quill-reasonQuill Researchreasoning5.0020.00100000

Models data dictionary

Check identifiers, dates, numeric fields, and missing values before writing a query or building a chart.

ColumnWhat it meansData typeMissing valuesKey or requirement
model_idStable model key.text0Required
model_nameFictional model name.text0Required
providerFictional provider, or Self-hosted.text0Required
tierflagship, fast, reasoning, or open_weights.text0Required
input_price_per_mtokUS dollars per million input tokens.decimal0Required
output_price_per_mtokUS dollars per million output tokens.decimal0Required
context_windowMaximum input tokens per call.integer0Required
View the full data dictionary

models

One row represents: One row per model.

ColumnWhat it meansData typeMissing valuesKey or requirement
model_idStable model key.text0Required
model_nameFictional model name.text0Required
providerFictional provider, or Self-hosted.text0Required
tierflagship, fast, reasoning, or open_weights.text0Required
input_price_per_mtokUS dollars per million input tokens.decimal0Required
output_price_per_mtokUS dollars per million output tokens.decimal0Required
context_windowMaximum input tokens per call.integer0Required

tools

One row represents: One row per tool.

ColumnWhat it meansData typeMissing valuesKey or requirement
tool_idStable tool key.text0Required
tool_nameFunction name.text0Required
categoryretrieval, data_lookup, action, code_execution, or document_processing.text0Required
requires_approval1 when a person must approve the call, otherwise 0.integer0Required

agents

One row represents: One row per agent.

ColumnWhat it meansData typeMissing valuesKey or requirement
agent_idStable agent key.text0Required
agent_nameAgent name.text0Required
use_casecustomer_support, data_analysis, code_review, sales_research, document_processing, or internal_help.text0Required
owning_teamTeam responsible for the agent.text0Required
launched_atDate the first version went live.date0Required

agent_versions

One row represents: One row per agent per deployed version.

ColumnWhat it meansData typeMissing valuesKey or requirement
agent_version_idStable version key.text0Required
agent_idAgent this version belongs to.text0Required
version_number1 for the first release, increasing by one.integer0Required
model_idDefault model for this version.text0Required
prompt_caching_enabled1 when prompt caching is on, otherwise 0.integer0Required
change_summaryWhat changed in this version.text0Required
deployed_atWhen the version went live (UTC).datetime0Required
retired_atWhen the next version replaced it; blank for the current version.datetime12Nullable

sessions

One row represents: One row per session.

ColumnWhat it meansData typeMissing valuesKey or requirement
session_idStable session key.text0Required
agent_version_idAgent version that was live when the session started.text0Required
channelchat, email, slack, api, or ide.text0Required
started_atSession start (UTC).datetime0Required
ended_atSession end (UTC).datetime0Required
outcomecompleted, escalated, abandoned, or failed.text0Required

llm_calls

One row represents: One row per model call.

ColumnWhat it meansData typeMissing valuesKey or requirement
call_idStable call key.text0Required
session_idSession the call belongs to.text0Required
call_numberOrder of the call within its session, starting at 1.integer0Required
model_idModel that served the call; can differ from the version default after a fallback.text0Required
started_atCall start (UTC).datetime0Required
latency_msTime to the complete response, in milliseconds.integer0Required
input_tokensInput tokens, including cached tokens.integer0Required
cached_input_tokensInput tokens read from the prompt cache.integer0Required
output_tokensOutput tokens.integer0Required
finish_reasonend_turn, tool_use, max_tokens, or error.text0Required
cost_usdCall cost in US dollars, rounded to six decimal places.decimal0Required

tool_calls

One row represents: One row per tool call.

ColumnWhat it meansData typeMissing valuesKey or requirement
tool_call_idStable tool call key.text0Required
call_idModel call that requested the tool.text0Required
tool_idTool that was called.text0Required
started_atTool call start (UTC).datetime0Required
latency_msTool run time in milliseconds.integer0Required
statussuccess, error, timeout, or denied.text0Required
error_typeReason for a call that did not succeed; blank on success.text124,780Nullable

evaluations

One row represents: One row per session per evaluator type.

ColumnWhat it meansData typeMissing valuesKey or requirement
evaluation_idStable evaluation key.text0Required
session_idSession that was scored.text0Required
evaluator_typellm_judge, human_review, or user_feedback.text0Required
score1 (poor) to 5 (excellent).integer0Required
passed1 when the score is 4 or 5, otherwise 0.integer0Required
failure_categoryMain problem when the session did not pass; blank when it passed.text9,664Nullable
evaluated_atWhen the score was recorded (UTC).datetime0Required

About the data

An AI agent leaves a trail: a session starts, the model is called several times, each call may run tools, and some sessions are scored afterwards. This synthetic dataset records that trail for twelve agents over calendar year 2025, with 30,000 sessions and the model calls, tool calls, and evaluations behind them.

Each agent also has a configuration history. A new version can change the default model, turn on prompt caching, or change the instructions, and every session points at the version that was live when it started. That lets you compare cost and outcomes before and after a release instead of averaging across the year.

What one row represents
Varies by source table
Data covers
Start: 2025-01-01 · End: 2025-12-31
Dataset version
1.0.0
Release published
2026-11-22

How the tables connect

Generated from the same relationship metadata as the table documentation.

Agent versionsagent_idAgentsagent_id

many-to-one

Agent versionsmodel_idModelsmodel_id

many-to-one

Sessionsagent_version_idAgent versionsagent_version_id

many-to-one

Llm callssession_idSessionssession_id

many-to-one

Llm callsmodel_idModelsmodel_id

many-to-one

Tool callscall_idLlm callscall_id

many-to-one

Tool callstool_idToolstool_id

many-to-one

Evaluationssession_idSessionssession_id

many-to-one

Start your first analysis

Open a download, inspect the table, and answer one question.

Start with Excel

Use the prepared one-row-per-session workbook to compare agents without writing joins.

Download Excel
File used
ai-agent-operations-v1.0.0-analysis.xlsx
Dataset version
1.0.0
File size
2.2 MB
  1. 01Open "ai-agent-operations-v1.0.0-analysis.xlsx" and select the "AI Agent Operations Dataset" sheet.
  2. 02Insert a PivotTable with agent_name in Rows and cost_usd and session_id (count) in Values.
  3. 03Add outcome to Columns to see how each agent's sessions end.

What you can build

Finished output from this release

Cost per completed session by agent

Model call cost is summed to the session first, then sessions are grouped by agent. Cost per completed session divides all of an agent's cost, including sessions that were escalated, abandoned, or failed, by its completed sessions.

AgentSessionsCompletedTotal cost (USD)Cost per completed session (USD)
pr-reviewer30002393797.930.3334
test-writer15001089295.660.2715
lead-researcher21001618169.040.1045
refund-assistant3300238195.060.0399
sql-analyst2700202159.020.0292
support-triage6000494653.30.0108
View the checked SQL
WITH session_cost AS (
  SELECT session_id, SUM(cost_usd) AS cost_usd
  FROM llm_calls
  GROUP BY session_id
)
SELECT a.agent_name,
       COUNT(*) AS Sessions,
       SUM(s.outcome = 'completed') AS Completed,
       ROUND(SUM(c.cost_usd), 2) AS Total cost (USD),
       ROUND(SUM(c.cost_usd) / SUM(s.outcome = 'completed'), 4) AS Cost per completed session (USD)
FROM sessions s
JOIN agent_versions v ON v.agent_version_id = s.agent_version_id
JOIN agents a ON a.agent_id = v.agent_id
JOIN session_cost c ON c.session_id = s.session_id
GROUP BY a.agent_name
ORDER BY total_cost_usd DESC;

Agent cost report

Report cost per session and per completed session for each agent by month. Show how much of the change comes from volume and how much from cost per session.

Release comparison

Pick one agent and compare each version with the one before it on completion, tool errors, and cost. State which release you would have rolled back, and why.

Tool reliability review

Rank tools by failure rate and find the week in which one tool's errors spiked. Separate tool errors from approvals a person denied.

Session fact model

Design one row per session with cost, tokens, tool errors, and evaluation results, built from the call-level tables without double counting.

Questions about this dataset

Is this real usage data from an AI product?

No. It is synthetic data for a fictional company. Model names, providers, and prices are invented, so the figures are for practice and are not benchmarks for any real model.

Does it include prompts or model responses?

No. Each model call records token counts, latency, finish reason, and cost, and each tool call records the tool, timing, and status. There is no prompt, response, or tool argument text.

How is the cost of a model call calculated?

Uncached input tokens and output tokens are charged at the model's price per million tokens in the models table, and cached input tokens at 10% of the input price. The result is rounded to six decimal places, and every call reconciles to that rule.

Can I compare agent versions?

Yes. agent_versions has one row per release with its dates, default model, and a summary of the change, and each session points at the version that was live when it started.

Source, license, and versions

This dataset models a fictional business. Every record was generated for practice.

Before you interpret the results

  • This is synthetic educational data generated to model a fictional business. It does not represent a real company or establish industry benchmarks.
  • Model names, providers, and prices are invented. They are not the prices of any real model.
  • Prompts, responses, and tool arguments are not included, only counts, timings, and outcomes.
  • Evaluations cover a sample of sessions, so pass rates describe the scored sample.
  • No PII is included.

Source, license, and version

Cite the version number when you share your work so anyone can rerun it on the same files.

Maintainer
Analytics Engineering
License
CC BY 4.0
Version
1.0.0
Published
2026-11-22T17:02:31.085Z

Cite this dataset

Analytics Engineering. (2026). AI Agent Operations Dataset (Version 1.0.0) [Data set]. https://www.analyticsengineering.com/datasets/ai-agent-operations.

Found an issue?

Send the dataset, version, and file you used, plus what you expected to see. A small example gets it fixed faster.

Report a dataset issue →

Keep learning

Cost lives on model calls, errors on tool calls, and quality on evaluations. Aggregate each to the session before joining them, or the totals multiply.

Practice exercises run in your browser on a smaller sample of this dataset.