Agentic Problem Frames Paper
Key Takeaways
- Five canonical agentic problem frames: Search, Planning, Execution, Reflection, Coordination
- Frame selection determines appropriate architecture — wrong frame → wrong agent design
- Most enterprise failures map to frame mismatch, not model capability gaps
- Cognitive depth should match problem complexity: shallow for routine, deep for novel
- Hybrid frames (Search+Execution, Planning+Reflection) needed for complex real-world tasks
Summary
This paper introduces a taxonomy of “agentic problem frames” — a structured way to classify what kind of cognitive work an agent task actually requires. The core argument: most agent system failures are not model failures but architecture failures stemming from applying the wrong problem frame.
The five canonical frames:
- Search: navigating a large information space to find relevant content
- Planning: decomposing a goal into a sequence of achievable sub-steps
- Execution: carrying out well-specified actions reliably with minimal deviation
- Reflection: evaluating outputs against criteria and revising iteratively
- Coordination: managing dependencies and communication across multiple agents or humans
The paper’s practical contribution is a decision tree for frame selection based on task characteristics: how well-specified is the goal? How large is the action space? How reversible are actions? The decision tree maps to architecture recommendations (single-agent vs. multi-agent, tool-use depth, human checkpoints).
Enterprise case study: a contract review workflow initially built as a Planning frame agent (generating review checklists) failed because the real bottleneck was inconsistent Execution. Reframing as Execution + Reflection reduced error rates by 73%.