New Trends in Harness Engineering: Deep Integration with Infrastructure

Introduction

In AI engineering, the boundary between "harness" (testing/evaluation frameworks) and "infra" (infrastructure) is becoming increasingly blurred. Recently, a technical expert pointed out on Twitter that many challenges in harness engineering are now best solved through tighter integration between harness and infra, citing async subagents and browser use as two examples, while announcing that these features are now included in MDA 0.10. This view reveals an important trend in AI system development: the convergence of toolchains is accelerating.

Why Do Harness and Infra Need to Be Integrated?

In traditional AI development, the harness typically refers to the framework used for testing and evaluating models or agents, while infra refers to the underlying infrastructure that supports operation, such as compute resources, networking, and storage. The two are often managed by different teams, resulting in high collaboration costs and difficulty in pinpointing problems. As AI agents grow more complex—especially those that need to interact with environments—the harness must be able to simulate real environments, while infra needs to provide flexible resource scheduling. The separation between the two leads to inconsistency between test results and production environments, thereby reducing development efficiency.

The expert noted that many challenges in harness engineering are now best addressed through tighter integration. This means that future harnesses may no longer be standalone tools, but rather deeply embedded into infra, sharing configuration, monitoring, and scheduling capabilities.

Agent in the AI context typically refers to a system capable of perceiving its environment, making decisions, and taking actions to accomplish specific goals, with autonomy and the ability to interact with the environment as its core characteristics. Unlike traditional single-pass inference models, agents often require multi-step planning, tool invocation, state maintenance, and even collaboration with other agents. This complexity means that testing and evaluating agents can no longer be limited to accuracy on static datasets; their behavior in dynamic, open environments must be examined. Therefore, the harness needs the ability to simulate environments, inject faults, and observe intermediate states—capabilities that are inseparable from the compute, networking, and storage resources provided by infra. Understanding this nature of agents helps clarify why the separation between harness and infra becomes an engineering bottleneck.

Integration Cases: Async Subagents and Browser Use

Async subagents and browser use are two typical examples. Async subagents allow an agent to execute tasks in the background while the main agent continues with other work, which requires the harness to simulate asynchronous behavior and verify results. Browser use involves an agent controlling a browser to interact with web pages, which requires the harness to provide a real browser environment and work closely with infra's networking and compute resources.

In MDA 0.10, these features are supported. MDA is likely an agent development framework (the specific name is not clarified in the original text), and its new version simplifies the testing and deployment of these complex scenarios by integrating harness with infra. For example, testing async subagents may require simulating latency and concurrency, while browser use requires managing the lifecycle of browser instances—both of which can be efficiently implemented through integration at the infra level.

Async subagents is an agent architecture pattern: the main agent delegates time-consuming or parallelizable subtasks to one or more subagents, which execute independently in the background. The main agent can continue processing other work without blocking and waiting, and finally aggregates the results. This pattern can significantly improve throughput and responsiveness, but it also introduces new testing challenges—verifying concurrency safety, timeout handling, result consistency, and failure retries. Traditional synchronous testing frameworks struggle to simulate realistic asynchronous timing and resource contention, so infra-level scheduling and monitoring capabilities must be leveraged to construct reproducible asynchronous scenarios. MDA 0.10's support for this pattern is achieved precisely by deeply combining the harness's assertion logic with infra's task orchestration.

Browser use refers to an agent controlling a real browser through programmatic interfaces to complete interactive tasks such as web navigation, form filling, and information extraction. It is an important capability for measuring an agent's practicality in open web environments. Its testing challenges include: dynamically changing web content, dependence on JavaScript rendering, and issues such as anti-scraping and login state management—making it difficult to cover real behavior with mocks alone. Incorporating browser use into the harness means the test environment needs to launch, isolate, and recycle browser instances on demand, while ensuring network access is consistent with the production environment. This is a typical scenario for harness-infra integration—browser instance lifecycle management, resource quotas, and network policies naturally fall under infra's responsibilities, while assertion and evaluation logic fall under the harness's responsibilities. Only through tight integration of the two can efficient and trustworthy testing be achieved.

Implications for AI Engineering Practice

This trend has important implications for AI engineering practice. First, when building AI agents, teams should prioritize a unified design of harness and infra to avoid the complexity of later integration. Second, when choosing frameworks, teams should pay attention to whether they support this kind of deep integration—for example, MDA 0.10 provides relevant features. Finally, developers need cross-domain knowledge, understanding both the logic of agents and the operation of infrastructure.

Although the original information is limited, we can infer that this integration will push AI agent development toward greater efficiency and reliability. In the future, we may see more frameworks building harness functionality into infra, creating an integrated development experience.

Conclusion

The integration of harness and infra is an important direction in the evolution of AI engineering. Through cases such as async subagents and browser use, we can see how this integration addresses practical challenges. The MDA 0.10 update is a concrete manifestation of this and deserves developers' attention. As the technology matures, this pattern is expected to become standard practice.