RESEARCH QUESTIONS / WORKING NOTES
What are we trying to learn?
The same practical question runs through the NOAERTH portfolio: can shared intelligence and reliable tools help a small team do meaningful work across many fields?
1. Can projects share what they learn?
PAIOS and other memory-oriented work explore continuity between sessions, tasks, and products. The challenge is keeping useful context while letting people correct it.
2. How much work should software do on its own?
AgentOS and Portfolio OS explore delegation, review, and safe handoffs. We want to learn which work benefits from automation—and where human decisions should remain central.
3. How do we know the work is good?
Builds, tests, and independent checks matter more than a system saying it has finished. We are developing clearer evidence for what was attempted, what passed, and what still needs work.
4. Can one approach serve very different projects?
Games, music, enterprise tools, and small software products give us different environments in which to ask the same question. The aim is to learn what transfers—and what does not.
How we approach it
For one detailed experimental direction, see the agent-coordination study specification and its descriptive scoring prototype.
Relevant work
AutoGen studies cooperating agents. FrugalGPT examines efficient model selection. SWE-bench asks how to evaluate coding tasks. Generative Agents explores memory and interactive behavior.