In harness scaling, the transferability of execution controls follows a model-distance hierarchy where the transferable artifact becomes progressively more abstract as architectural and task distance increase:

1. **Frozen Transfer Within a Model Family**: Closely related model generations or variants (such as GPT-5.5 to GPT-5.6) share recurring execution failure modes. This allows an exact, frozen control profile—including concrete states, activation policies, routing logic, pre-commit checks, and repair routines—to transfer successfully without any target-model modifications.
2. **Adapted Transfer Across Model Providers**: Crossing provider boundaries (such as OpenAI to DeepSeek) introduces distinct base-agent behavioral variations, causing direct frozen transfer of concrete profiles to fail or cause slight regressions. However, transfer succeeds at an architectural and methodological level: the generic runtime, high-level runbook structure, golden rules, and failure-analysis loop remain fully reusable, requiring only lightweight adaptation of concrete practices to the target model's residual failure distribution.
3. **Task-Side Generalization Across Distributions**: Across heterogeneous task families, transfer operates at the level of principles and mechanisms—specifically the methodology for identifying, locating, and enforcing sparse, consequential execution boundaries.

Model-Distance Transfer Hierarchy in Agent Harnesses

Cross-provider harness adaptation is the process of adjusting concrete execution practices in an agent runbook when transferring an execution harness across different model providers while maintaining the generic runtime, runbook structure, and development principles intact.

Because autonomous agents built on models from different providers retain significant behavioral autonomy and exhibit distinct residual failure distributions, directly applying a frozen control profile engineered for one provider can fail to improve or slightly regress performance. Rather than engineering a harness from scratch, cross-provider adaptation reuses:
- The underlying state runtime and transition enforcement semantics;
- The high-level runbook graph and phase-level boundaries;
- Domain-independent golden rules (such as preferring minimal reusable control and semantic routing over task-identity overfitting);
- The failure-driven postmortem optimization loop.

Adaptation is strictly confined to analyzing residual failure traces on the target model to update provider-specific practices, state-local prompts, and targeted verification checks, enabling lower-cost models to achieve frontier performance with minimal API adaptation expenditure.

Cross-Provider Harness Adaptation

Harness scaling is the systematic enhancement of the execution and control layer surrounding an autonomous agent without changing its underlying model weights. Rather than replacing model scaling, harness scaling investigates how much of an agent's latent capability can be converted into completed, reliable work by improving runtime mechanisms that maintain durable state, constrain valid execution, verify progress before transitions, and recover from errors.

Harness Scaling

Harness scaling investigates whether external execution controls can convert a model's latent capabilities into reliable work through three progressively stronger empirical evaluations:

1. **Fixed-Model Improvement (Within-Model Lift)**: Assesses whether a refined execution harness can improve the end-to-end task completion rate of a fixed model without altering its underlying weights.
2. **Cross-Generation Transfer (Frozen Transfer)**: Assesses whether an execution control profile developed on an earlier model generation can transfer directly to a newer, more capable model without retuning or modifying control logic.
3. **Beyond-Benchmark Generalization (Held-Out Task Generalization)**: Assesses whether the resulting control principles and runtime mechanisms can improve performance on completely different, held-out task distributions beyond the benchmark on which they were originally developed.

Three Empirical Tests of Harness Scaling

University of Michigan - Ann Arbor

Ch.2 Provider Transfer and Model Hierarchies - Autonomous Agent Control Planes: State Scaffolding and Resilient Execution @ University of Michigan - Ann Arbor

Model-Distance Hierarchy and Cross-Provider Transfer - Autonomous Agent Control Planes: State Scaffolding and Resilient Execution @ University of Michigan - Ann Arbor

Learn Before

Related

Learn After