TESTING
Cold-context review harness
Tooling that spawns a reviewer agent with zero implementation context and diffs its findings against human review. Measuring where independence actually pays.
LAB
Experiments in Progress
Unfinished by design. Discarded experiments stay visible. Knowing when to stop is part of the method.
PROTOTYPING
Handover compressor
Turns a long agent session into a minimal context document a fresh session can resume from. The hard part is deciding what to forget.
INVESTIGATING
Answer-engine visibility
How programmatic local pages surface in AI answer engines versus classic search. Early, mostly measurement design.
PAUSED
Visual regression via browser agent
An agent operating a browser for regression checks. Paused: a screenshot-diff script does 90% of it at 5% of the cost.
DISCARDED
Fully autonomous dependency upgrades
Discarded after two near-misses. Upgrades that pass tests can still change behaviour tests never covered. Now agent-drafted, human-merged.
SHIPPED
Environment doctor
A single script that validates a full local setup and prints exactly what is missing. Boring, and used every week.