Skip to content

Investigate adaptive escalation policy for Plan-Execute orchestration #360

@HemVadgama

Description

@HemVadgama

Motivation

The AssetOpsBench paper notes that Plan-Execute agents can exhibit limited sensitivity to task complexity and may over-plan relatively simple objectives.

I would like to investigate whether an adaptive escalation policy could improve efficiency while maintaining answer quality.

Hypothesis

Simple tasks should often be solvable by a central planner with minimal agent/tool involvement.

More complex, ambiguous, or high-risk tasks may benefit from escalation to additional reasoning, verification, or specialist workflows.

Initial Experiment

Compare baseline Plan-Execute against a lightweight escalation wrapper that:

  1. Performs an initial assessment.
  2. Determines whether escalation is necessary.
  3. Executes additional verification/reasoning only when triggered.

Metrics

  • Benchmark score
  • Tool calls
  • Agent calls
  • Latency
  • Token usage
  • Escalation frequency

Status

Initial investigation and architecture review.

Metadata

Metadata

Assignees

Labels

No labels
No labels

Type

No type
No fields configured for issues without a type.

Projects

No projects

Milestone

No milestone

Relationships

None yet

Development

No branches or pull requests

Issue actions