Production Agent
Prompt quality stops mattering when memory, tools, retries and tenancy disagree. The system is the product.
- LangGraph
- RAG
- Temporal
- Observability
THREAD 01 · SYSTEMS THAT REMEMBER
What makes an agent
usefully present?
Abstract. I build AI products around a simple problem: intelligence is not enough. A useful system must preserve context, take bounded action, expose uncertainty, and recover when reality disagrees.
- Author
- Huang Guoqiang
- Location
- Hangzhou, CN
- Practice
- AI-native product engineering
- memory
- action
- recovery
- human boundary
Selected systems as
an ongoing study.
MemoHome
Care interfaces require domain boundaries, offline recovery and consent semantics before visual “accessibility” can be honest.
- Multi-client
- DDD
- Offline-first
- Audit
Oneness AI
Generative creativity becomes editable only when assets, queues, cancellation and compensation share one state model.
- Next.js
- Postgres
- BullMQ
- Providers
Unimage
Scene facts and spatial relationships survive model changes better than decorative adjectives.
- Audit
- Spatial facts
- Compile
- Validate
Me Observe
Personal memory stays useful when facts, inference and human confirmation remain visibly different layers.
- Evidence tiers
- Local files
- Git
- Human review
NOTES ON THE OBSERVER
F1, new music, old machines, quiet paintings.
These are not unrelated hobbies. They are ways of watching systems change over time: a race lap, a new song, a horizon, the feedback between a tool and a hand.