The test harness
Human auditor
Reviews flagged findings · decides what "good" means next
Context
Code
Design
Requirements
User stories
Bug history
Release notes
Support tickets
always live · never static
Rules · the harness
Intent
Quality criteria
Boundaries
Skepticism
Audits
what "good" means for this product
Agent
> observe
> compare against intent
> flag anything suspicious
probe
observe
Product under test
STEP 01Login as trial user
STEP 02Create workspace
STEP 03Invite teammate
STEP 04Run first job
STEP 05Verify billing
real product · real state · real data
out
Verified quality
✓ Pass
✕ Fail
⚑ Flag
every run · every time
How it works: Seven live context streams feed into the rules layer, which defines what "good" means for that specific product. The agent operates inside those rules, probing the product and flagging anything suspicious. A human auditor reviews the findings and tightens the rules, making the harness smarter with each cycle.