Generative AI
Document processing isn't one problem — it's three
Harness engineering is an active and developing concept in AI-assisted software engineering. A harness is essentially the support we give a coding agent to help it do its job, so harness engineering is all the work we do around that to make that possible.
In the first half of 2026 the industry has been paying significant attention to skills, markdown files we give to agents to help them complete specific tasks. While skills have undoubtedly been crucial in driving the concept of harness engineering, there is a powerful and underdiscussed component of the agent harness that deserves attention: sensors.
In a recent video, we took a closer look at how sensors can help us better harness AI agents in software development. Watch below.
Feed-forward vs. feedback mechanisms
When we talk about an outer harness (the environment surrounding the agent), we can divide it into two categories:
Skills and guardrails (feed-forward): The instructions and constraints we provide to give context to the generation process. They include coding conventions, principles, and reference documentation.
Sensors (feedback): These are tools that observe what the agent actually produced. They provide a loop for the agent to self-correct before a human even looks at the code.
In a perfect world we would only need feed-forward because the agent would get it right the first time. Since we don’t live in that world, sensors are vital for delivering software safely and sustainably.
The power of computational feedback
Sensors can be inferential (the agent must interpret the signal) or computational (deterministic tools like linters and test runners).
Successful teams are moving away from basic prompts where you repeatedly ask the agent to write a test and hope it’s diligent enough to follow your instructions, toward deterministic constraints that provide guarantees. LLMs judgement is great if you’re operating an exploratory area with fuzzy rules, but once you want something more objective and consistent, adopting a formal, unambiguous and deterministic tool gives you more assurance.
Real-world experimentation: Sensors in action
In a deep-dive experiment on a TypeScript data dashboard, we tested the impact of running a coding agent with and without a suite of sensors, including:
Static analysis: ESLint and Semgrep.
Structural checks: Dependency Cruiser to enforce module boundaries.
Test quality: Coverage reports and mutation testing.
We found that when the coding agent was equipped with sensors that give it feedback on code quality, it was able to improve quality over time e.g. increasing test coverage.
How sensors help agentic engineering
The key takeaway is that sensors maintain standards. They make warnings manageable, help us identify the key signals in the noise and direct us to spend energy in the right places.
A dashboard of green or red sensors can help a developer quickly assess a codebase’s bill of health and give us an indication of where we need to focus our attention and reinvest in the harness. It’s this work that ultimately distinguishes sustainable development from chaotic vibe coding.
Sensors also support coding agents. When we give them both inferential and computational ways of evaluating their work, we enable them to act on feedback without human intervention. This lowers the burden on the human coder and allows them to focus more on design intent.
It’s worth also bearing in mind that harness engineering isn't about total automation; it's really about situational awareness for the developer. Reading a massive diff from an AI agent is exhausting; humans should sit on top of a higher-abstraction steering loop.
Looking ahead: Harness templates
In the future, we will move away from building harnesses from scratch for every project. Instead, we might use harness templates tailored to specific application types — like a data dashboard template or a CRUD business service template — that come pre-configured with the ideal guides and sensors.
Ultimately, sensors allow us to use coding agents more effectively and ensure consistency and quality. They offer us the ability to leverage advanced quality assurance techniques that were previously too high-friction for humans to maintain manually.
