AI Agent Retro Kit

Last Updated on October 10, 2026 | By AI-Augmented Scrum | Sprint Retrospective

Bring your sprint’s agent data. Classify what went wrong as a requirement, context, or model problem, and leave with a focused retro agenda and experiments for next sprint. Everything runs in your browser.

Inspect the System, Not the Agent

An agent does not carry lessons from one sprint to the next. Your configuration does: the requirements it was given, the context it could read, the permissions it had, and the review that followed. That is what a retrospective can actually change, so that is what this kit asks about.

The Requirement-or-Model Test

Re-run a failed task with a clearer requirement and the missing context. If the result improves, the cause was the input, and a human process needs fixing. If it still fails, you have found a real capability limit. Enter the outcome below and the kit sorts each defect for you.

The Kit

1. Sprint data

Pull these from your repository, CI, and agent logs. Leave a field blank if you don't track it.

Agent PRs opened

Agent PRs merged

Avg review rounds per merged PR

Tasks with retry or rewrite loops

Fabricated references caught

Invented APIs, packages, or file paths

Cost this sprint

Tokens or currency; use one unit

Cost last sprint (optional)

Same unit as above

2. Classify agent-related defects

Add the two or three worst outcomes. For each, answer four questions. The kit applies the requirement-or-model test: if a clearer requirement and better context fix the result, the cause was the input, not the model.

How the Kit Classifies a Defect

Each defect lands in one of four causes, based on your answers. The thresholds that raise signals from your sprint data are starting heuristics, not benchmarks; tune them to your team after a few sprints.

Separately, any defect that should have been caught in review or CI is flagged as a review gap.
Cause	How the kit decides	Typical experiment
Requirement	The requirement was vague or incomplete, and a clearer re-run improved the result (or the re-run was not tried).	Add acceptance criteria and a worked example to agent-bound stories.
Context	The agent lacked needed context, or the result improved on re-run even though the requirement was clear.	Write the missing knowledge into the repository instruction and documentation files.
Model limit	The task still fails with a clear requirement and sufficient context.	Narrow the scope, add a human step, or route the work to a person.
Undetermined	Requirement and context look fine and the re-run has not been tried.	Re-run before the next retro so it can be classified.

Run the Whole Retro

The kit gives you evidence and a starting agenda. The article explains how to run the conversation, and the other tools cover the rest of the sprint.

How to Run the Retro All Tools Read the Guide