University of Michigan - Ann Arbor

Control drift is an optimization defect that emerges when iteratively adapting an agent harness across a heterogeneous benchmark suite. When developers or hyper-agents attempt to resolve failures on individual tasks one at a time, localized rules and checks introduced to fix a specific task can impose unnecessary procedural overhead, over-constrain the model's action space, or inadvertently degrade performance and cause regressions on unrelated tasks across the benchmark.

Control Drift in Multi-Task Harness Optimization

In multi-task harness optimization, what iterative practice triggers control drift?

Control drift is a defect caused by fine-tuning shifts in the model's underlying neural network weights.

Identify two distinct operational consequences that localized rules and checks inflict on an agent when control drift occurs.

Sparse routing is an architectural pattern designed to allow a single, shared runbook to govern diverse, heterogeneous tasks without inducing control drift. Rather than encoding exhaustive, per-task procedures into the global workflow, sparse routing organizes control into:

1. **A Thin Shared Runbook**: Maintains a lightweight, general state backbone common across all tasks.
2. **Semantic Triggering**: Activates specialized pathways and policies derived from observable task semantics rather than static task identifiers.
3. **Late State-Local Checks**: Enforces specialized transition checks and constraints only when current task context and workspace artifacts explicitly demonstrate elevated downstream failure risk.

This structure preserves broad agent autonomy on routine tasks while selectively enforcing stringent controls only where specific failure modes are likely to occur.

Learn Before

Related