The Human in the Loop: When AI Orchestrates and You Execute
It's mostly endorphins' fault, but during a forty-five minute tempo run on a Tuesday morning, the thought struck me with such clarity that I nearly stopped mid-stride: I am the human in the loop.
Not in the sense we typically discuss it—as the wise overseer keeping wayward algorithms in check, the guardian ensuring AI doesn't run amok. No, this was something else entirely. I was literally the only biological component in an otherwise fully automated system. The workout plan was generated by Anthropic's Claude. The interface that displayed my targets was built by Google's Gemini. The post-run analysis would be performed by OpenAI's GPT. And somewhere in the orchestration layer, yet another model was coordinating all of this, switching seamlessly between providers based on the task at hand.
My role? Run the prescribed intervals. Be the servo motor in an optimization loop I'd designed but no longer actively controlled.
When your training plan comes from a committee of AIs
Until recently, "human in the loop" meant something quite different. It suggested human authority, human oversight, human control. A reassuring phrase implying that despite all the automation, a person remained ultimately in charge. What I experienced that morning was a profound inversion: the loop itself was AI, and I was simply the wetware executing its instructions.
The Context Collapse of Agency
What strikes me most about this reversal is how naturally it happened. I built Mera—my personal virtual coaching platform—to solve a specific problem: preparing physically and mentally for a trek to Mera Peak (6,476m). The system needed to understand my baseline fitness through Strava integration, interpret my heart rate zones, comprehend my goals and their horizons, and dynamically adjust training blocks based on how my body responded to the workload.
The architecture emerged from practical necessity. Different language models excel at different tasks. Anthropic's models proved remarkably thoughtful at constructing training periodization—understanding the relationship between base building, intensity progression, and recovery cycles. OpenAI's GPT excelled at analyzing completed workouts, parsing heart rate variability and lactate threshold data to provide actionable feedback. Gemini Flash, running the conversational interface, offered quick responses to questions like "Should I run tomorrow if my legs feel heavy?"
But sneaky things happened while I was busy appreciating this division of labor. The system developed a kind of collective intelligence. Not through some emergent property or secret model communication, but through something simpler and more profound: each AI specialized in its domain, and I became the integration point—the biological middleware executing their combined directives.
The Evolution of Virtual Coaching
That may explain why this particular Tuesday's tempo run felt so revelatory. The workout itself was unremarkable: 10-minute warmup, three 8-minute intervals at tempo pace (around 80-85% max heart rate), 2-minute recovery jogs between intervals, 10-minute cooldown. Standard training block for building aerobic capacity.
But consider the orchestration:
Planning Phase (Claude): Analyzed my recent training history, upcoming trek timeline, current fitness level from Strava data, and recovery patterns from my training journal. Generated a 16-week periodized plan with specific workouts, intensity targets, and adaptation triggers.
Display Layer (Gemini + Astro): Rendered the day's workout on my phone with real-time heart rate zone indicators, pace targets, and interval timers. The interface updated dynamically as I ran, comparing actual performance to prescribed targets.
Execution Phase (Me): Ran the intervals. Focused on maintaining effort in the prescribed zones. Provided the biological processing power the system lacked—the actual muscle contractions, the cardiovascular response, the mental discipline to sustain uncomfortable effort.
Analysis Phase (OpenAI): Post-run, ingested my Strava data, compared actual heart rates to targets, evaluated pace consistency across intervals, analyzed heart rate recovery between efforts, and generated insights about my current fitness adaptations.
Adaptation Phase (The whole system): Used my workout notes ("intervals felt easier this week, especially the third one") combined with the quantitative analysis to inform next week's training block adjustments.
Look with me at what happened here: four different large language models, five if you count the orchestration layer, coordinated to guide a single human body through a scientifically designed workout. The human provided nothing to this system except execution and feedback. The intelligence—all of it—was artificial.
When Execution Becomes the Human Domain
All of this computational coordination swims around inside a larger cultural shift we're still struggling to name. We built AI to augment human intelligence, to handle tasks that were tedious or complex or required processing at scale. Somewhere along the way, for certain domains at least, the relationship inverted.
Virtual coaching offers a particularly clear example because the value chain is so transparent. The intelligence required to design a training plan that balances physiological adaptation, injury prevention, goal specificity, and individual recovery needs is considerable. It requires understanding exercise physiology, periodization theory, individual variability, and the interplay between different training stimuli.
Traditionally, this intelligence came from human coaches—professionals who spent years learning these principles, accumulating experience with different athletes, developing intuition about what works. The limitation was always scale and personalization. A coach could work with only so many athletes, and even then, the level of individualization was constrained by time and attention.
Large language models, trained on vast repositories of exercise science literature, coaching methodologies, and training data, can generate personalized plans at scale. But they can't run the workouts. They can't provide the biological feedback loop. They need a human in the system—not to oversee them, but to execute the plan they've designed.
The Abstraction That Changed Everything
What makes this work—what transforms it from an interesting experiment into a genuinely useful system—is something I've used in other projects: LLM abstraction. The idea is simple: instead of hardcoding which model handles which task, create a layer that can route tasks to different providers based on criteria like cost, latency, capability, or performance.
In practice, this means I can change the model responsible for training plan generation from Anthropic to OpenAI to Google in seconds through an admin interface. If I discover that GPT-4's workout analysis isn't as nuanced as Claude's, I swap them. If Gemini releases a model that's faster at conversational responses, I route the chat interface to it.
This architectural pattern does something subtle and powerful: it makes the human role in the system even more purely about execution. I'm not debugging model behavior or compensating for AI limitations. I'm not manually combining outputs from different systems. The abstraction layer handles orchestration. The specialized models handle their domains. I run.
The philosophical implication is worth sitting with. In knowledge work, we've become accustomed to thinking of AI as a tool—powerful, certainly, but ultimately subordinate to human direction. "The AI suggests, the human decides" goes the reassuring mantra. But what happens when the AI's suggestions are consistently better than your untrained intuition? When its analysis is more rigorous, its pattern recognition more accurate, its consistency more reliable?
You stop being the decision-maker and become the executor. The loop doesn't disappear—it just changes who's providing the intelligence and who's providing the action.
Beyond the Tempo Run: The Broader Pattern
That's where this Tuesday morning revelation connects to something larger. Virtual coaching is just one domain where this inversion is happening, but the pattern is emerging everywhere:
Software Development: GitHub Copilot and Claude Code suggest implementations, developers execute keystrokes. The intelligence increasingly comes from the model; the human provides judgment about which suggestion to accept and the physical act of coding.
Creative Writing: Tools like Claude can generate drafts, analyze structure, suggest improvements. Writers increasingly curate and edit rather than compose from scratch.
Medical Diagnosis: AI systems trained on millions of cases can identify patterns in imaging or lab results that individual doctors might miss. The physician becomes the executor of the AI's insights, the human interface with the patient.
Financial Analysis: Algorithmic trading systems process market data and execute trades faster than humans can perceive. Traders monitor and adjust parameters rather than making individual trading decisions.
In each case, the traditional hierarchy inverts. The AI doesn't assist with human decisions; humans execute AI-directed actions. We're not teaching AI to be more human; we're learning to be better executors of AI intelligence.
The Preservation of Biological Necessity
But still more themes present themselves. If humans become primarily executors in AI-orchestrated systems, what remains distinctively human about the loop?
The answer, I think, lies in what AI fundamentally cannot do: inhabit a biological body. My virtual coaching system can design perfect training plans, but it cannot run them. It can analyze workout data, but it cannot feel the burning in my quads during the third tempo interval. It can prescribe rest days, but it cannot experience the deep fatigue that signals genuine overtraining.
This biological grounding matters more than it might seem. The feedback I provide to the system—"intervals felt easier this week"—comes from embodied knowledge that no amount of sensor data fully captures. The system learns not just from my heart rate and pace, but from my subjective experience translated into words.
This creates an interesting partnership. The AI provides intelligence the human lacks (comprehensive knowledge of exercise science, pattern recognition across thousands of training progressions, tireless consistency in plan execution). The human provides capabilities the AI lacks (biological execution, subjective experience, embodied wisdom about what "hard but sustainable" actually feels like).
Neither component is subordinate. Each is essential. The loop requires both.
The Future of Collaborative Execution
What this Tuesday's tempo run suggests is that we need new language for human-AI collaboration. "Human in the loop" carries connotations of human oversight and control that don't match the reality of these emerging systems. "AI assistant" suggests subservience that doesn't reflect how these tools actually function at scale.
Perhaps what we're discovering is something more symmetrical: collaborative execution. The AI provides intelligence—planning, analysis, optimization. The human provides embodiment—action, feedback, integration. Neither is the boss; both are necessary.
This interpretation transforms how we might think about AI deployment across domains. Instead of asking "How do we keep humans in control?" we might ask "What can only humans provide?" and "What can only AI provide?" The answer isn't always "humans provide oversight." Sometimes it's "humans provide execution" or "humans provide embodied feedback" or "humans provide contextual judgment."
In my case, building a virtual coaching system that uses multiple large language models wasn't about replacing human coaches. It was about recognizing that the intelligence required for training design is now readily available through AI, while the biological execution remains necessarily human. The system works because each component handles what it does best.
Twenty weeks from now, when I'm hopefully standing on Mera Peak at 6,476 meters, the achievement will be simultaneously mine and not mine. I'll have provided the biological effort—the thousands of training hours, the discipline, the execution of countless workouts like Tuesday's tempo run. But the intelligence that guided that effort? That came from the loop. The AI loop, with one human component.
The magic, as I'm discovering, isn't in maintaining control over AI systems. It's in recognizing when the most valuable thing a human can provide is simply showing up and doing the work the machines designed but cannot execute themselves.
After all, somebody has to run the intervals.
Want to build your own AI-orchestrated training system? Mera is currently in personal use but demonstrates the power of multi-model coordination for virtual coaching. The principles apply to any domain where human execution remains essential but AI intelligence can provide superior planning and analysis.
