I spend a large part of my day working alongside AI. Not experimenting from a distance, but building with it, thinking with it, and trying to move faster because of it. And recently I’ve started noticing something uncomfortable.
The thing slowing me down is no longer the technology. It’s me.
AI can go all day. I can’t.
There’s an assumption that once you learn how to prompt well, everything gets easier. In practice, I’m finding the opposite. The more capable AI becomes, the heavier the mental load feels.
Every interaction demands context switching. I need to decide what I want, how to phrase it, how much background to include, how to validate the output, and whether I trust it enough to move forward. Multiply that across coding, architecture decisions, writing, planning, and experimentation, and it adds up quickly.
I can manage one or two deep AI workflows at a time if I start early. Beyond that, my brain starts pushing back. Not because the AI is underperforming, but because the responsibility of directing it never switches off. AI doesn’t get tired. I do.
The trust problem we don’t talk about
If I’m honest, part of the fatigue comes down to trust.
When I code on my own, I know exactly why something exists. When AI generates code, even when I know it’s good, there’s a layer of uncertainty from my side. I still feel accountable for understanding every decision before allowing it into the system. I am still the one that must “report” if it breaks, so how can I not ensure I understand it before letting it go?
That accountability creates friction.
I slow down. I double-check. I re-read. I refine prompts. I try to maintain control over something that can clearly move faster than I can reason about it. The irony is hard to miss: as AI becomes more powerful, I spend more mental energy managing it.
I keep looking for ways to rely on it with greater confidence and spend less time second-guessing each step. But letting go of that control isn’t simple. I am trying different frameworks for validation of output — I haven’t found the golden goose yet. I am sure though that checking AI’s work is the wrong way to go about it. The bottleneck isn’t intelligence — AI has enough of that. It’s ownership.
We still treat AI like a tool waiting for instructions
Most workflows today assume the human must initiate everything.
- Open a chat.
- Write a prompt.
- Wait.
- Review.
- Redirect.
In the beginning, this felt powerful. Almost revolutionary. I would move through the steps and, at the end, have something tangible to review and decide whether I agreed with it.
That sense of control felt good. Now it feels different.
I rarely review every single thing AI produces in the way I once did. Not because I trust it blindly, but because I understand it better. I’ve learned how to work with it. I’ve built guardrails. I’ve added checks and balances. I design the environment so the outputs are directionally sound before they appear.
But that doesn’t mean I won’t review it. It means the review has shifted from inspecting every brick to assessing the structure.
The process still works. But it doesn’t scale.
I can only manage so many parallel conversations before they start blending into noise. Meanwhile, AI remains idle unless I initiate something. If I stop working, it stops working for me.
And that’s the limitation.
We’re still operating in a model where intelligence waits for instruction. Where momentum depends entirely on human activation. There has to be a better way. That’s where I think the real shift happens. Not smarter prompts. Different interaction.
How this might evolve
I don’t believe the long-term solution is simply “get better at prompting”. Prompting feels like an early interface, not a permanent one. If AI continues to improve, it can’t remain dependent on humans typing instructions all day. The system itself has to evolve.
Think back to the early days of mobile phones. The breakthrough was simple and thrilling: you could make a call from anywhere. That alone felt revolutionary. We genuinely believed the future of the device was about being able to talk on the phone without being tied to a wall.
Today, calling is one of the least interesting things our phones do.
Our primary interaction isn’t voice at all. It’s messaging, navigation, payments, photography, automation. In some of the best moments, we aren’t even holding the device. It unlocks as we approach. It surfaces information before we search. It routes us without us asking. It sits in the background, anticipating rather than waiting.
The original interface wasn’t wrong. It was just early.
Prompting may be the same. Right now, we are still “making calls”. Opening a chat. Typing instructions. Waiting for a response.
But if AI keeps advancing, interaction won’t stay anchored to manual input. It will move toward presence instead of initiation. Toward systems that understand intent and context without requiring constant translation.
Just like the phone stopped being primarily about calling, AI may stop being primarily about prompting.
Autonomous context instead of constant prompts
One of the biggest mental drains is repeatedly explaining context. A more sustainable model might involve AI continuously observing systems rather than waiting for explicit input.
Imagine AI reviewing commit histories, analytics dashboards, performance metrics, and architecture patterns on its own. Instead of me initiating every task, it proposes actions, with reasoning attached. It wouldn’t replace decision-making. It would reduce the effort required to begin thinking. That shift alone would remove a significant portion of the mental load.
I can see evolution happening in different ways here.
1. Ambient collaboration instead of active control
Right now, working with AI feels like managing a conversation thread. I suspect it moves toward something more ambient. AI embedded inside the workflow. Monitoring changes. Surfacing insights at the right moment.
Instead of asking, “What should I do next?” it flags when something drifts from intent or when an opportunity appears. Less prompting. More reviewing.
2. Parallel agents instead of a single stream
Another limitation is that I still interact with AI as if it were one entity. One task at a time. One thread at a time. But AI is capable of parallel processing way beyond what we use it for most of the time.
I can imagine defining direction once, then allowing multiple agents to move simultaneously. One exploring architecture options. Another running experiments. Another analysing risks. My role shifts from operator to orchestrator. That feels like the only way humans keep pace with where AI is heading.
3. AI initiating work instead of waiting for ideas
One of the more surprising realisations is how often AI waits for me to think of something first. Yet AI already sees patterns I don’t. If it’s analysing data, behaviour, and system changes continuously, why should it wait for me to articulate a problem?
The future may involve AI bringing forward ideas before I consciously recognise them. Not replacing strategy, but expanding what’s visible. That would change the dynamic entirely. Humans would no longer be the sole source of momentum.
4. Moving beyond text prompts
Text prompts themselves are probably temporary. They require precision and context. Both of those require energy. Over time, interaction may shift toward intent rather than instruction. Visual models. System diagrams. Behavioural goals. Constraints defined once instead of restated endlessly.
The core skill may not be prompt engineering at all. It may be designing systems so AI understands direction without constant translation.
Why this matters
Some of the scenarios I mentioned earlier are already starting to happen. AI is becoming more embedded. More proactive. More context aware.
I’m not a futurist. I’m not claiming to see what’s coming with certainty. I can only observe the trajectory and make an informed guess.
But the direction feels clear.
I don’t believe humans are becoming obsolete. But we may be becoming the slowest moving part of a rapidly accelerating system. And that creates tension.
AI can explore continuously. Humans have finite attention. AI scales with compute. Humans scale with focus. AI operates without pause. Humans need space to think.
The solution won’t be asking people to work harder or faster. It will come from redesigning collaboration so that thinking doesn’t feel like manual labour.
Right now, working with AI feels powerful, but heavy. I can see where it’s going. I can also feel the strain of being the one holding everything together.
Maybe the next evolution of AI isn’t about becoming more intelligent. Maybe it’s about becoming easier for humans to live with.
A guess: ten years from now, I’m still a lead software engineer
The title hasn’t changed. The work has.
My day doesn’t begin with coffee and a blank chat window. There’s no ritual of typing the first prompt, no careful phrasing to coax momentum into motion.
Instead, I open a system overview.
It feels less like opening a tool and more like stepping onto a balcony overlooking a living city. While I slept, the system was awake.
Several AI agents have been active through the night. One monitored performance regressions, quietly mapping subtle deviations before they became visible to users. Another explored alternative architecture patterns for a feature we’ve been circling for weeks. A third ran controlled experiments against staging traffic, probing edge cases no one had time to simulate manually. A fourth reviewed our security posture and tracked dependency drift across services.
Nothing dramatic. Just steady progress.
I don’t start by asking, “What should I work on?” Instead, I’m presented with a briefing.
Two architectural trade-offs are flagged for review. The first leans toward short-term delivery speed. The second favours long-term modularity. Both are modelled. Both are viable.
There’s an anomaly in user behaviour. A subtle shift in onboarding flow completion rates. It may signal friction. Or it may signal an unmet need.
There’s a proposed refactor, complete with quantified risk, rollback plan, and projected impact on maintainability.
And there’s a performance trend that won’t hurt us today, or tomorrow, but will become a real issue in three weeks if ignored.
Each item comes with reasoning. Evidence. Simulations. Second-order effects mapped out in ways that would have taken me days to assemble on my own.
My role isn’t to invent options from scratch anymore. It’s to evaluate direction.
Mid-morning, I join a design session. Ten years ago, this would have been a debate built on whiteboard sketches and intuition. Now, we explore live system models. When someone proposes increasing throughput at the API layer, we see cost curves shift instantly. Latency projections adjust in real time. Complexity indexes nudge upward as hidden dependencies surface.
We’re not arguing about guesses. We’re shaping constraints.
In the afternoon, I don’t assign tickets. I define intent.
- Optimise onboarding latency without increasing infrastructure cost.
- Increase resilience of this service under regional failure.
- Reduce cognitive load in this module for future engineers.
The agents take that intent and decompose it into parallel streams of work. One drafts implementation paths. Another runs simulations. Another stress tests failure scenarios. Another generates documentation and opens pull requests, each with embedded reasoning explaining why decisions were made.
Pull requests still exist. But they read differently now.
I review diffs less for syntax and more for coherence. Less for whether the code compiles and more for whether the system still feels clean. I look for creeping complexity. For hidden coupling. For trade-offs that look efficient today but expensive tomorrow.
My questions have changed. Is this aligned with product strategy? Are we introducing invisible complexity? Is this elegant, or merely fast?
By late afternoon, I realise something that would have felt impossible years earlier: progress hasn’t depended on how many hours I spent typing.
Work has continued in parallel. Experiments have run. Models have improved. Risks have been mapped. Proposals have matured.
The system hasn’t waited for me to think of everything first.
And I’m not obsolete. If anything, I’m more necessary.
But I’m no longer the engine trying to generate every idea, explore every branch, and hold every detail in my head.
I’m the one setting direction. The one deciding what matters. The one accountable for the shape of the system, not the speed of the keystrokes.
Ten years ago, I was managing conversations with AI. Now, I’m collaborating with momentum.
I am Ironman.
100% Human Written — 0 Agents were harmed in the creation of this post