University of Michigan - Ann Arbor

Multi-task harness abstraction is a two-tiered optimization workflow designed to synthesize durable procedural runbooks across broad task families:

1. **Task-Level Execution & Proposal**: Individual task agents execute training instances and propose localized procedural changes directly from their observed failure traces.
2. **Hyper-Agent Evaluation**: A more capable supervisory hyper-agent evaluates candidate proposals across multiple instances, evaluating their generality, safety, and cross-task compatibility.
3. **Reconciliation & Freezing**: The hyper-agent reconciles validated lessons into a shared family-level runbook and removes idiosyncratic task workarounds before freezing the profile for held-out evaluation.

This division of labor prevents singular instance failures from triggering over-specialized family rules while automating the synthesis of persistent procedural memory.

Multi-Task Harness Abstraction via Hyper-Agent

In the multi-task harness abstraction workflow, what is the primary responsibility of the supervisory hyper-agent during the evaluation phase?

In multi-task harness abstraction, individual task agents freeze the shared family-level runbook before submitting their failure traces to the hyper-agent.

Why does multi-task harness abstraction employ a two-tiered division of labor between task agents and a supervisory hyper-agent rather than allowing task agents to update the family runbook directly?

Learn Before

Related