Thoughts on Working with AI

Recently, I have spent a great deal of time building an agentic analytics platform. The experience has taught me more than technical lessons. It has also raised deeper questions about how humans work with AI—and how humans should work with one another in the age of AI.

In my work environment, each person may have multiple development environments and several AI agents working in parallel. AI can produce code and documents at remarkable speed. I serve as a key designer and contributor while also helping the project succeed: shaping the architecture, defining milestones, coordinating responsibilities, reviewing designs and code, implementing critical components, and maintaining the overall quality of the system.

This way of working has exposed several challenges. People bring different strengths to the project, and AI can help them contribute beyond their primary areas of expertise. However, the resulting code may be difficult to understand, validate, and integrate into the wider system. At the same time, I need to understand their pull requests while continuing to deliver my own work. The result is often cognitive overload: AI can generate material far faster than a human can carefully understand and validate it.

Looking back, the experience has strengthened several beliefs about working with AI.

1. Architecture and design matter more than execution

Our role is increasingly to shape the architecture and design rather than perform every implementation step ourselves. Even detailed work needs a clear structure, workflow, and set of boundaries.

Humans must understand and clearly articulate the goal, then design a credible path toward it. AI can help refine the plan and accelerate execution, but it should not determine the goal on our behalf. Faster implementation cannot compensate for unclear intent or weak architecture.

As AI takes on more implementation work, people should focus first on defining and reviewing goals, requirements, architecture, interfaces, constraints, and acceptance criteria. A clear design document gives humans and AI a shared contract before code multiplies. The resulting code still needs automated checks and focused human review—especially in high-risk areas—but our greatest leverage comes from validating intent and design before implementation begins.

2. Build systems with self-correcting loops

AI can execute individual tasks faster than we can. Our job is to think beyond each task and build a system that can continue to produce reliable results. That system includes both the product users receive and the development process used to build it. Both need feedback and self-correction.

A good system does more than produce the right result once. It continues to behave reliably as inputs, users, and conditions change. It makes its behavior understandable, detects when it is going wrong, limits the impact of failure, and either corrects itself or gives people enough information and control to intervene. This is especially important for AI systems, whose behavior can be probabilistic and difficult to predict. I think about a good system as a connected set of capabilities:

1. Correctness: Define what good output means and the constraints the system must satisfy.
2. Observability: Make the system's behavior, decisions, and failures visible.
3. Recoverability: Provide ways to contain failures, roll back changes, and restore a known-good state.
4. Adaptability: Incorporate new information and respond to changing conditions without losing control.
5. Safety: Establish boundaries that prevent unacceptable actions and protect people, data, and systems.
6. Testability: Make expected behavior verifiable through repeatable tests and evaluations.
7. Feedback: Feed production signals, test results, and human judgment back into the system so that it can improve or trigger correction.

Together, these capabilities form a self-correcting loop: define the target, observe deviations, contain failures, adapt safely, verify behavior, and feed what we learn into the next iteration.

The development process needs the same loop. We should not ask only whether a task is finished; we should ask whether our way of working can repeatedly produce reliable results. AI can help challenge assumptions, generate tests, monitor outcomes, identify regressions, and carry lessons into the next iteration. If AI expands implementation capacity without also expanding our ability to verify and correct the work, we will simply create more issues at greater speed.

3. Collaboration is becoming coordination among human–AI units

When a group works together without shared principles, coordination becomes increasingly difficult. AI amplifies this problem because every person can now produce changes at a much higher rate. In effect, each person and their AI agents become a small development unit that must coordinate with other human–AI units.

This creates a new working relationship. Each person should take responsibility for the quality of their own scope before sharing it with others. The goal cannot be to deliver as quickly as possible while leaving validation and integration to someone else.

People who performed well before the AI era often become even more effective with AI because the underlying methods remain valuable: define the process, establish principles, design the system, and take ownership of quality. AI amplifies what we can do, but sound engineering judgment determines whether that greater capacity leads to good outcomes.

4. Deep expertise and interpersonal skills both matter

Complex projects require many people to work together, and every important area needs someone who understands it deeply enough to protect its quality. AI can help people cross unfamiliar boundaries, especially when experts have already created strong frameworks and reusable components. But AI does not eliminate the need for expertise.

Building an AI product is partly like creating building blocks and partly like assembling those blocks into something useful. Experts design reliable blocks, interfaces, and constraints. The broader team can then combine them safely and efficiently. Without that foundation, apparent speed often creates hidden fragility.

Interpersonal skills have always mattered, but AI makes them even more important. When everyone can produce more work faster, unclear assumptions and poor coordination spread faster too. Valuable collaborators combine deep expertise with sound judgment, intellectual humility, and the ability to communicate and learn from others. They explain their reasoning, welcome review, and change course when evidence challenges either AI-generated output or their own conclusions.

They also take responsibility for how their work affects the wider system and bring a sense of craftsmanship to it. When visibility or credit matters more than the quality of the outcome, AI can amplify the resulting pressure and hidden risks across the team.

5. Privacy and security cannot be postponed

As development accelerates, risk can also grow. People have less time to understand what has been generated, what data it touches, and which assumptions it makes. When a privacy or security incident occurs, the response must be fast—but speed alone is not enough.

If we rely on AI without understanding the system ourselves, we may not even know whether the problem has truly been fixed. Human understanding remains the final line of accountability.

Protecting the human mind

One person can operate many AI agents but still has only one brain. As we greedily ask more agents to work for us, we must protect that brain. It is like being surrounded by delicious food and trying to eat everything without considering the limits of the stomach.

Mental health and cognitive health have never been more important. AI can surround us with an endless stream of code, documents, decisions, and notifications. Cognitive overload may become one of the defining problems of AI-assisted work. We need to learn how to let the mind recover.

Although this is an age obsessed with efficiency, we should deliberately allow our minds to do nothing: rest, walk in a forest, play without a screen, and sleep well. Rest is not the opposite of productive thinking; it is part of the system that makes good thinking possible.

Human beings did not come into this world to become machines. When AI enters a project, it is easy to become completely outcome-oriented and forget the people doing the work. We should care for our colleagues and organize work in a way that gives people something beyond a salary: opportunities to learn, a sense of craftsmanship, mutual respect, and meaningful contribution.

AI should help us build better systems without making us less human. If our pursuit of efficiency harms everyone involved, then the system is not truly successful—no matter how quickly it delivers.