# Agent workflow experiment agreement

Experiment ID/version: __________
Owner: __________
Written before run on: __________

## Decision
What specific pilot or product decision will this evidence inform?
__________

## Task and eligible cases
User: __________
Trigger/request: __________
Reviewed artifact: __________
One trial starts when: __________
One trial ends when: __________
Deadline: __________
Eligible cases and sampling method: __________
Planned exclusions, applied equally to both conditions: __________

## Comparison
| Detail | Baseline | Candidate |
|---|---|---|
| Workflow | | |
| Source snapshot/version | | |
| Model and configuration | | |
| Prompt/instructions | | |
| Tool/server version | | |
| Human assistance allowed | | |
| Browser/API support, if relevant | | |

Differences other than the connection: __________
Execution order / how learning effects are reduced: __________

## Hypothesis
For **[eligible tasks]**, **[candidate]** will improve **[primary outcome]** relative to **[baseline]**, while meeting **[critical conditions]**.

## Measures
Primary: reviewed acceptance by the deadline, using the attached rubric.
Secondary: elapsed minutes including review and corrections; review minutes; correction count; direct cost where recorded.
Diagnostic only: tool calls, retrieval latency, token counts, first-response time.

## Missing and unresolved records
- Unresolved by deadline: not accepted for the primary outcome.
- Unknown completion time: `null`, never zero or an imputed average.
- Timing comparison: use complete pairs only and label their case count.
- Incomplete cost records: report cost metric as unavailable.
- Exclusions: retain case ID and reason in the log.

Any changes to this policy before running: __________

## Review and decision
Reviewer(s): __________
How disagreements are resolved: __________
Conditions for a limited pilot: __________
Conditions requiring revision: __________
Conditions for stopping: __________
Review date: __________
