↓ Skip to main content
  1. writings/
  2. AI-Assisted Data Engineering/

Room to Breathe

·6 mins·
Table of Contents

Introduction
#

Working fast with agents makes it easy to sprint until you break. Leaving room to breathe between pushes is what keeps the work sustainable and keeps the judgment sharp enough to be worth accelerating in the first place.

The applied chapters showed where the practice pays off in real data work. The last part is about lasting, because none of that matters if you burn out before the practice compounds. AI changed the rhythm of work in a way that works against rest, and naming that change is the first step to protecting yourself from it. This chapter is the case for building the pause back in on purpose.

Pace over Sprint
#

Before AI, the supply of one more improvement was limited by how fast I could think one up. Coming up with the next idea took real effort, so a natural pause was built into every iteration, a moment to breathe between the last improvement and the next. AI removed that pause. The next idea is now free and instant, and there is always another variant waiting the moment I finish the last one, so stopping starts to feel like leaving performance on the table even when stopping is clearly the right call.

The trap underneath this is what I call phantom throughput. It feels like I can ship far more than before, because the model can produce far more than before, but my real rate of shipping work I actually understand has not moved an inch. The model multiplied the supply of possible work, not my capacity to deliver it, and mistaking the first for the second is how the sprint mentality takes hold. The dashboard of activity looks incredible while the amount of genuinely owned, verified work stays exactly where it was.

A sustainable pace is not slower, it is deliberate about where the limit really is. The work still has to pass through a human who reads the change, understands it, confirms it does what it claims, and owns it once it is live. Pace over sprint means setting your rhythm to that real rate rather than to the model’s output rate, because the model’s rate is a number you can never actually hit and exhausting yourself chasing it produces worse work, not more of it.

Space Between Pushes
#

The deeper reason to leave space has nothing to do with energy, it is about correctness. A new solution needs time to actually run in production before I pile the next change on top of it. Until it has handled real traffic and real edge cases for a while, I am not improving a known quantity, I am guessing on top of a guess. Giving it room to run is how I get a reliable signal about whether it even needs the next iteration, and surprisingly often the real answer is that it does not, at least not yet.

The solution also needs time to settle in my own head. If I jump straight into reworking something the moment it ships, I am iterating on a thing I do not fully understand yet, and that is how I lose the thread of my own system. A little distance is what turns “it works” into “I know exactly why it works,” and that second state is the one I need to be in before I am qualified to change it again. The space between pushes is where understanding catches up to output.

This is where less genuinely can be more. Stopping at something that is good enough and fully understood beats an endless squeeze that leaves me worn out and holding a system I can no longer reason about clearly. The restraint is not laziness, it is what keeps the work durable. Shipping fewer, better-understood changes and letting them breathe consistently beats shipping a constant stream of marginal ones I cannot keep straight, both for the system and for the person maintaining it.

Protecting Attention
#

Attention is the scarce resource that AI throughput drains, and protecting it requires deliberate time away from the screen. A screen rewards continuous output and trains you to produce on demand, and the assistant amplifies that by making the next prompt always available, so the session never has to end. The pull to ask for one more variant feels like collaboration, which makes stopping feel rude, as though you were walking out on a colleague mid-sentence. The assistant is not going to tell you to stop. That has to be your job.

The counterweight is what I think of as brain defrag. When I am away from the screen, moving, outside, doing something physical, my attention is not idle, it is unscheduled. Ideas that were pinned under urgent get to float, two half-formed thoughts from last week bump into each other, and a problem I could not crack at the keyboard often resolves itself once my mind changes channel. It feels like the old disk defragmenters sliding scattered fragments into order, and the machine runs faster afterward precisely because it spent time reorganizing instead of producing.

So I treat rest as a required input to the work, not a reward for finishing it. Recovery is part of the pipeline, the same way sleep is, and I schedule for it deliberately rather than hoping it fits in the gaps. I know not everyone has the same options for how to unplug, and the point is not the specific activity. The point is protected mental slack, whether that is a walk around the block, a shower with no podcast, or ten minutes staring at the ceiling without a phone. Good work depends on being rested enough to tell a good idea from a loud one, and that discernment is the first thing to go when the screen never lets you breathe.

Putting It Into Practice
#

  1. Set your pace to your real rate of owned, verified work, not to the model’s rate of output.
  2. Watch for phantom throughput, the feeling of shipping more when your understood output has not actually grown.
  3. Let a new solution run in production long enough to tell you whether it even needs the next iteration.
  4. Give a change time to settle in your head until “it works” becomes “I know why it works” before reworking it.
  5. Choose fewer, well-understood changes over a constant stream of marginal ones you cannot keep straight.
  6. Schedule deliberate time away from the screen, and treat that recovery as part of the pipeline, not a luxury.
  7. Accept that the assistant will never tell you to stop, so make stopping your own deliberate decision.

Related

Patterned Modeling, Human Judgment

·6 mins
Introduction # The best AI-assisted modeling work happens where the pattern is strong and the judgment boundary is clear. Data Vault 2.0 is a useful example here because hubs, links, satellites, and effective satellites are repeatable enough for an agent to scaffold, while the decisions that make the model correct still belong to the human.

Declarative Pipelines Are Easier to Assist

·6 mins
Introduction # AI agents work best when the system they are helping with is legible. Declarative pipelines make the desired result visible in one place, which gives both the agent and the human reviewer a better surface to reason over than a scattered chain of tasks, schedules, and merge scripts.

Teaching the Agent Where History Belongs

·6 mins
Introduction # The first AI move around history is not asking the agent to build a history table. It is asking the agent to argue against needing one. Stored history is stateful, expensive to recover, and easy to misuse, so the assisted workflow has to begin with restraint rather than generation.

Generating Tests and Documentation

·7 mins
Introduction # Tests and documentation are the most neglected parts of data engineering because they are tedious to write by hand. An agent that reads a model, infers the business logic, and generates the matching definitions gives you the rigor without the manual overhead, as long as you still review the result.

Avoiding AI Brain Fry

·7 mins
Introduction # AI brain fry comes from too many parallel prompts, rapid context switching, and weak stopping conditions. The workload feels productive while the cognitive overhead compounds, so sustaining the practice means managing yourself as carefully as you manage the agent.

When Models Take Liberties

·7 mins
Introduction # As models get more capable they get more confident, and that confidence makes them take liberties, such as pushing to a primary branch or running a query nobody asked for. The last chapter described the guardrail pattern. This one explains why those guardrails need to exist before the model gets a chance to improvise.

Speed with Lane Discipline

·7 mins
Introduction # Guardrails are not brakes, they are the lane markers that let me drive fast because I know exactly where the road edges are. The goal is never to slow the agent down on the work that is safe, only to remove the handful of ways it can do something I cannot undo.

Wiring Tools with MCP

·7 mins
Introduction # MCP is how the agent reaches the tools where the work really lives, including issue trackers, repositories, and databases. Wired well, it turns a messy brain dump into a structured issue that becomes the contract the agent reads while it implements.