University of Michigan - Ann Arbor

Instance-dependent harness demands refer to the principle that optimal agent execution scaffolds vary not only across broader task domains, but across individual problem instances. Different computational tasks inherently require distinct structural priors:

- **Wide-search tasks** benefit from parallel evidence exploration workflows.
- **Terminal and command-line tasks** favor lean, serial ReAct execution loops.
- **Deep-research tasks** require structured working memory to manage and synthesize retrieved evidence.
- **Repository-level coding tasks** require filesystem mediation to organize code patches, unit tests, execution traces, and repository states.

Because of these disparate demands, optimizing a single ahead-of-time scaffold across heterogeneous tasks is cumbersome and necessitates searching an expansive design space with no guarantee of matching the next task's structure.

Instance-Dependent Harness Demands

According to the principle of instance-dependent harness demands, which structural prior is specifically required for repository-level coding tasks?

Why is optimizing a single ahead-of-time scaffold across heterogeneous tasks problematic?

Just-in-Time (JIT) agent harness construction is an execution paradigm where a task-specific operational scaffold is synthesized dynamically on the spot at inference time rather than precompiled ahead of time. Under the JIT paradigm, a trained generator receives the target task specification, operational protocol, available capability registry, and prior harness context to assemble an adapted harness instance. By tailoring memory management, planning representations, and action loops to the exact problem at hand, JIT construction amortizes harness optimization and shifts agent design from repeated offline search to runtime synthesis and evolutionary revision.

Learn Before

Related