Research Questions | Evan Ma

01 / Research Questions

Q01

How should an agent leave work that can be accepted and handed off?

Current hypothesis: A completion claim should travel with evidence, boundaries, validation, and a next owner.
How I am testing it: Test the protocol through SACP, Agent Flight Recorder, and real execution receipts.
Still unresolved: How can acceptance become independent of an agent's own honesty?

Q02

What should survive context compression and window changes?

Current hypothesis: Long-term memory should prioritize decisions, facts, evidence, risks, and next actions.
How I am testing it: Observe information loss through handoffs, compaction, worklogs, and Solo AI Company OS.
Still unresolved: How should information be routed among context, project memory, and long-term knowledge?

Q03

How can people and AI compound judgment, not only task throughput?

Current hypothesis: Human goals, taste, reflection, and private evaluation determine whether AI becomes compounding learning.
How I am testing it: Track changes in judgment through delivery, writing, workflow reviews, and learning systems.
Still unresolved: How can we show human judgment improved instead of merely increasing AI output?