Understand what your agent work costs.
Bring run and call records, confirm what they mean, and see why a customer's agent became more expensive. One period gives task economics. Two comparable periods explain a change.
Working browser-based analysis of supported execution exports. Saving in this browser is optional; hosted workspaces are not available yet.
Sample data · fictional
Explore sample data
Clearform (fictional) sells a document-review agent to insurers on fixed monthly contracts. Two months of generated runs and calls. Every figure on the results page is computed from those records.
Templates
The sample dataset in the supported format. Match these columns and the importer will accept your export.
Your records
Upload execution records
Two CSV files in the supported format, or a workspace file exported from this tool. No install, no contract details, no questionnaire.
Where your records go
This page is served over HTTPS from our host. Your files are read in this tab and only the supported columns are kept as normalized records. This flow sends no records to the server and sets no cookie. Identifiers and metadata can still be sensitive, so use exports without prompts, documents or contact details.
Unchecked: records stay in memory for this tab and are gone on refresh or when the tab closes. Export a workspace file from the results page to keep your work.
Supported format and templates
Two CSV files. This is the one format the importer accepts; other exports are not parsed. Download the templates above and match the columns.
- Two CSV files: runs and calls. This is the only format the importer accepts; arbitrary exports are not parsed.
- A task is one unit of work. A retry is a run whose parent is another run and whose relationship is retry. A subagent is a child run whose relationship is subagent.
- Costs of failed, abandoned and unknown runs are always kept.
- Money is measured only when cost_usd is present. Otherwise it is estimated from tokens at dated list prices and labeled as such.
- No prompts, completions or documents. Ids, timestamps, labels, counts and amounts only.
runs.csv columns
| column | required | also accepted | meaning |
|---|---|---|---|
| run_id | yes | runid, id | Unique id of one agent run. |
| started_at | yes | start_time, timestamp, created_at | ISO 8601 instant, e.g. 2026-08-12T14:03:00Z. Periods are calendar months in UTC. |
| customer_id | no | customer, tenant_id, account_id | Your own customer key. Without it, results stay company-level. |
| task_id | no | task, job_id | Groups runs into one unit of work. Without it, each root run is a task. |
| parent_run_id | no | parent_id, parent | The run this one belongs to, for retries and subagents. |
| relationship | no | relation, run_kind, child_kind | retry or subagent. A parent link alone cannot tell them apart. |
| workflow | no | workflow_name, agent | Which agent or workflow ran. |
| workflow_version | no | version, prompt_version | Version label of that workflow. |
| outcome | no | result, task_status | Your outcome label. You map labels to success, failure, abandoned or unknown on the next step. |
calls.csv columns
| column | required | also accepted | meaning |
|---|---|---|---|
| run_id | yes | runid | The run this call belongs to. |
| model | no | model_id | Model name. Needed for list-price estimates when no cost is given. |
| input_tokens | no | prompt_tokens | Measured usage. |
| output_tokens | no | completion_tokens | Measured usage. |
| cached_input_tokens | no | cached_tokens | Cached prompt tokens, if your provider reports them. |
| cost_usd | no | cost, amount_usd | Your own recorded cost for the call. When present it is used as measured money instead of an estimate. |
| call_id | no | request_id, id | Optional id for drill-down. |
| occurred_at | no | timestamp, created_at | When the call happened. Defaults to the run start for rate lookup. |
| provider | no | vendor | openai, anthropic, … |
| status | no | http_status | Serving status of the call. This is not task success. |