Model updating is the ability to revise an agent’s understanding of a problem and its plan when newly arriving evidence contradicts its initial hypothesis. section 10.1.3
Iteration alone is not enough: an agent may continue following its original loop or plan even when the situation no longer fits. Updating turns contradiction into a reason to replan. section 10.1.1
A practical mental model is:
- Start with a hypothesis about what the user needs. section 10.1.3
- Monitor evaluation signals for evidence that conflicts with the hypothesis. section 10.1.1
- Revise the plan rather than executing the original sequence unchanged. section 10.1.3
A contradictory user scenario
Suppose an agent selects a standard resolution path during decomposition, then receives user information that contradicts that path. Without model updating, it follows the standard plan anyway. With updating, the planning module enters replanning mode and constructs a revised path. section 10.1.1
Table 10.1 Agent cognitive capabilities
| Capability |
Present |
Absent |
Supporting primitive |
| Task decomposition |
Agent invents novel subtask sequences |
Agent treats complex queries as flat lookups |
CoT, planning |
| Dependency reasoning |
Agent sequences steps while respecting dependencies |
Agent executes steps in arbitrary order |
ToT, sequential thinking |
| Compositional tool use |
Agent chains tools in novel configurations |
Agent maps queries to single tool calls |
ReAct |
| Model updating |
Agent revises plan when evidence contradicts it |
Agent follows initial plan regardless |
Reflexion, replanning |
Compare model updating with other agent capabilities, including the absent behavior of following an initial plan regardless of evidence and the supporting primitives Reflexion and replanning. section 10.1.3
