Software Engineering Research in the Agent Era: Complexity, Knowledge, and Harness Engineering

The piece opens with a quote from Demis Hassabis: intelligence ultimately reduces to pattern recognition in massive noisy data, and intuition is an algorithm science can uncover.

Shift in Research Focus

A referenced article notes that software engineering research is moving from a human-centric core — how to develop and maintain software efficiently — to an agent-centric paradigm dubbed Harness Engineering . The fundamental goal remains the same: achieve high-quality, efficient development and business success, but now the question is how to use agents to reach it.

Core Problem: Software Complexity

The author argues that the essential challenge for agent-driven development is identical to the human-driven one: software complexity , citing the "First Law of Software Engineering: Complexity Law." From this root, four high-level research questions emerge:

Complexity characterization: What are the dimensions of software complexity? How can we measure them and evaluate where agents perform poorly?

Machine-readable software knowledge: How to represent requirements, architecture, and decisions so agents can read, write, and execute them? What schema, expression, and organization methods are needed?

Agent team collaboration: Humans align via documents, meetings, and chat. How do agents align when they depend on external tools? How do humans align their own use of diverse agents?

Adapting traditional SE theory, methods, and tools: How must existing practices evolve for agent-based workflows?

Decomposition into Concrete Sub-questions

The author further breaks down the first two high-level questions:

What specific dimensions define software complexity characterization (referencing another article on "hybrid and scattered" complexity)?

How to automatically compute complexity from those dimensions?

How to generate evaluation benchmarks from real projects (e.g., AOSP, Linux kernel) using the defined complexity metrics?

What constitutes software knowledge? How should each type be recorded, organized, continuously extracted, and precipitated across requirements, design, coding, and testing? This ties into agent observability and memory management.

Guiding Principles for Research Investment

The author proposes two filters to avoid low-return work:

Avoid directions easily absorbed by foundation models: Generic agent frameworks, tool calling, and general software engineering knowledge are already well-handled by major labs.

Avoid shallow topics: As agent frameworks stabilize, rebuilding another framework yields diminishing returns. Core valuable modules (memory, search engines) are gradually separating from the harness layer.

The article concludes that many deep problems remain to be excavated, but researchers should focus on areas where they can contribute unique depth rather than competing on commoditized infrastructure.