Agentic Problem Frames
Definition
An engineering framework adapted from classical software engineering Problem Frames theory, providing structured approaches to designing reliable AI agents. It introduces Agentic Job Descriptions (AJD) and the Act-Verify-Refine (AVR) loop.
一种改编自经典软件工程问题框架理论的工程框架，提供了设计可靠智能体的结构化方法。它引入了智能体职位描述（AJD）和“行动-验证-优化”（AVR）循环。
Details
Three Fatal Risks in Current Agent Development
- Unbounded scope — agents acting beyond their authority
- Knowledge gaps — agents filling ambiguity with hallucinated assumptions
- Open-loop execution — no verification of outcomes
1. **无界范围** —— 智能体越权操作
2. **知识缺口** —— 智能体通过产生幻觉式的臆测来填补模糊性
3. **开环执行** —— 缺乏对结果的验证
Agentic Job Description (AJD)
Like a human job description, an AJD defines:
- Jurisdictional boundaries — what the agent is allowed to do
- Operational contexts — environments and constraints
- Knowledge anchors — domain knowledge sources
- Evaluation criteria — how to measure success
- Feedback mechanisms — how to learn and improve
正如人类的职位描述一样，AJD 定义了：- **管辖边界** —— 智能体被允许执行的操作 - **运行语境** —— 环境与约束条件 - **知识锚点** —— 领域知识来源 - **评估标准** —— 如何衡量成功 - **反馈机制** —— 如何学习与改进
Act-Verify-Refine (AVR) Loop
- Act — execute according to dynamically generated specification
- Verify — check outcome through callbacks and user confirmations
- Refine — update knowledge for next iteration
1. **执行** — 根据动态生成的规格说明执行
2. **验证** — 通过回调和用户确认检查结果
3. **优化** — 更新知识以用于下一次迭代
This shifts from static one-shot specification to progressive convergence.
这从静态的一次性规范转向了**渐进式收敛**。
Key Insight
The framework treats the agent as a black box and instead structures its interactions with the environment.
该框架将智能体视为黑盒，转而对其与环境的交互进行结构化。
Connections
- Related to: Harness, Multi-Agent Architectures
- Mentioned in: APF Paper
- 相关：[[ai-agent-architecture/concepts/harness|Harness]]，[[ai-agent-architecture/concepts/multi-agent-architectures|多智能体架构]] - 提及于：[[ai-agent-architecture/sources/agentic-problem-frames-paper|APF 论文]]