This post explores the core philosophy of proactive agent design through a 'problem transfer' (问题转移) mechanism, illustrated with two architecture diagrams:
Key Idea: Problem Transfer Over Task Abstraction
Traditional AI/agent pipelines tend to follow: receive user question → abstract into a generic task → execute → return results. This introduces semantic drift: the boss asks A, the system abstracts it into B, and what gets executed is effectively a different problem.
The proposed philosophy instead:
- Receives the question and transfers it as-is, preserving its original form
- Rejects over-abstraction — the AI should not create demand, "optimize" the user's question, or "intelligently" reinterpret it
- Casts the AI as a faithful problem courier, not an unauthorized problem rewriter
- Every task transfer must retain the full context of the original question
- Prohibit "smart summarization," "automatic optimization," and other operations that deform the problem
- Understanding may happen at the execution layer, but the transfer layer must stay faithful
- Historical memory is not a task-production source; each task starts from zero
- Context may serve as *reference*, but not as *basis*
- Tasks are effectively stateless, akin to pure functions in functional programming — avoiding task drift, priority confusion, and overloaded inference
- Before execution, audit: (1) who raised the question and what was their intent; (2) the transmission chain and any intermediate deformation; (3) the background and relevant context
- The output of this audit is refinement, not reconstruction
- The agent's "proactivity" should manifest in auditing tasks, not inventing them
Three Guiding Principles
1. Never forget the original intent
2. Travel light
3. Vet three generations before proceeding
Practical Value
1. Explainability — original questions are preserved; every transfer is traceable and auditable 2. Controllability — users always see their original question and can intervene before execution 3. Efficiency — lightweight reasoning; fewer wrong-task executions and rework 4. Safety — prevents AI-generated pseudo-demands and task drift
Open Challenges
1. Granularity: overly vague questions ("make me something") may be unexecutable if transferred verbatim; a balance between fidelity and sufficiency is needed 2. Cost: auditing three generations on every transfer adds latency; verification must be efficient 3. Learning: full statelessness conflicts with experience accumulation; the author suggests distinguishing *task context* (discard) from *capability context* (retain)
Conclusion
The philosophy is both retro (a return to simple, traceable system design) and avant-garde (a challenge to mainstream AI paradigms that emphasize intent understanding and autonomous decision-making). Its viability depends on engineering implementation and product conviction.