Approval Gates, 3 Watchers, and AI Fitness Coaching

Approval Gates, 3 Watchers, and AI Fitness Coaching

Why constrained autonomy beats fully automated coaching when the body is noisy, the data are incomplete, and the cost of a bad call compounds fast.

Justin Harris’s coaching logs show a recurring pattern: when clients drifted into constipation, he did not hand control to a bot or let a model “optimize” around the issue; he adjusted food, water, and timing first, because distension in prep was tied to dry gut content and backed-up output, not a mysterious failure mode. That is a small example of a larger mechanism: approval-gated control. In AI fitness coaching, the best architecture is not “autonomous agent with access to everything,” but “agent drafts, human approves,” because the error costs of premature automation rise faster than the convenience benefits. My thesis is simple and falsifiable: as coaching decisions get closer to body composition, sleep, fatigue, and recovery tradeoffs, autonomy should shrink, not expand.

The problem with “just automate it”

The hype version of AI coaching assumes more data automatically means better decisions. In practice, more data often means more ways to be confidently wrong. The corpus here keeps pointing to the same reality: coaches are not just reading numbers; they are interpreting context.

Take Harry Sims. He reported two daily caffeine exposures, 150 mg with black coffee at meal 1 and another 150 mg from a white monster with meal 2, plus a sleep schedule that varied by workday and commute demands. The useful detail is not the caffeine total in isolation. It is that he volunteered an exportable WHOOP record, noted no evidence of apnea, and said he woke feeling well rested. A machine can count caffeine. A coach has to decide whether that caffeine is a problem in this person, on this week, inside this prep.

That is the central limitation of full automation: the model sees a pattern; the coach sees a pattern with a provenance chain. Did the data come from a device? A self-report? A transient travel week? A food swap? The answer changes the decision.

Why approval gates belong in the middle, not the edges

Approval-gated workflows are not a slower version of autonomy. They are a different risk design.

In a good coaching stack, AI should do three things well:

1. summarize the week,
2. flag anomalies,
3. propose a draft next step.

It should not, by default, execute the next step.

That distinction matters because coaching decisions are often reversible in theory but not in practice. When Justin wrote that he would avoid pushing prep hard for more than 16 weeks, that was not a generic “less is more” slogan. It was a constraint on exposure to hard dieting because the duration itself is part of the stress load. Similarly, his note to keep a client in a good position at the current body fat and dose level was not a promise of linear progress. It was an argument for preserving margin while the system still had room.

AI tools are good at pushing toward the mean. Coaching is often about avoiding the mean outcome for this one person, this week, in this state. Approval gates force the system to stop at the point where human judgment is still required.

The failure mode is confident overreach

The dangerous version of fitness automation is not a cartoon robot ordering impossible workouts. It is subtler: a model quietly over-optimizes one metric because it cannot appreciate second-order effects.

Justin’s comments around prep distension are a good case study. He linked worsening distension to backing up during prep, less water than usual on travel, and the gut drying out faster as the diet got harder. He also noted that clients with narcotic abuse tended to have the worst distension because constipation is one of the main side effects. That is not a recommendation to do anything pharmaceutical; it is a warning about causal chains. What looks like “the plan stopped working” may actually be a hydration, GI, or adherence problem.

A fully autonomous coaching agent can easily miss that distinction because it lacks a safety culture. It wants to recommend. It does not naturally want to ask: what else changed? travel? water intake? stress? bowel regularity? The best approval gate does exactly that. It slows the system just enough to ask for the missing causal link before an action gets locked in.

Constrained autonomy beats open-ended agency

There is a useful way to think about this: AI should be allowed to be broad in observation and narrow in action.

Broad observation means the system can ingest check-ins, weights, sleep logs, training notes, and symptom reports. Narrow action means it can only make bounded recommendations within a preset playbook unless a human approves escalation. That is not a bug. It is the product.

Why? Because fitness coaching has a lot of “known unknowns.”

• A client says they slept 8:30 on average, but that average hides commuting days and routine shifts.
• A client says back pumps are worse, but the actual driver may be training load, hydration, or the current phase of the diet.
• A coach sees lower back stubbornness or depletion and has to balance visible leanness against the rebound that comes later.

In each case, the output should be a draft and a confidence note, not an executed change. The more the system is allowed to change food, training, supplements, or recovery targets on its own, the more likely it is to “solve” the visible metric while harming the hidden one.

What approval-gated automation looks like in practice

For coaches, the useful question is not whether to use AI. It is where to put the gate.

A practical structure looks like this:

• Step 1: Ingest. AI pulls in check-ins, device summaries, training logs, and coach notes.
• Step 2: Triage. AI highlights only what changed materially: weight trend, sleep disruption, GI complaints, unusual fatigue, or a mismatch between reported and observed state.
• Step 3: Draft. AI proposes one to three adjustments, each tied to a stated reason.
• Step 4: Approve. The coach accepts, edits, or rejects the draft.
• Step 5: Escalate only with human sign-off. Anything that affects the whole plan, changes risk materially, or touches ambiguous signals stays gated.

That workflow is slower than a fully agentic system, but it is faster than an overconfident one, because it avoids cleanup. Most coaching damage is not caused by ignorance alone. It is caused by premature certainty.

The falsifiable test

If approval-gated automation is the right model, then the strongest AI coaching systems will not be the most autonomous ones. They will be the ones that reduce coach workload without increasing the number of unreviewed plan changes.

That can be tested.

Track how often the model makes a recommendation, how often a human approves it, and how often the final decision is later reversed or corrected. If full autonomy were superior, higher agent permission should correlate with fewer corrections and better continuity. If constrained autonomy is superior, the best-performing systems will show the opposite pattern: lots of drafts, high human override value, and fewer bad downstream surprises.

That is the standard coaches should care about. Not whether the AI sounds smart. Not whether it can talk like a coach. Whether it helps the coach stay decisive without letting the machine outrun the evidence.

Bottom line

AI fitness coaching should not be designed around trust in autonomous output. It should be designed around trust in constrained process. Let the machine summarize, compare, and propose. Let the human approve the action. In a domain where small mistakes compound, autonomy is not the default virtue; restraint is.

Sources Used

• raw/_consumed/2026-05-31/kahunas-export/2026-05-31-w19-24m/clients/david_lamartina___members-tlssnsjthkmnhfqcscszce25acz_vhdm_x2_xdlpx_i.json
• raw/_consumed/2026-05-31/kahunas-export/2026-05-31-w13-18m/clients/skip_hill___members-b7s-_d0oqqz4a1tadlvuztb0y_vzfqlr9c29tpakweg.json
• raw/_consumed/2026-05-31/kahunas-export/2026-05-31-w19-24m/transcripts/alex_goracy___members-m1vvgmmbnhazzipv5wqmwwhgjuyotawsompy4kzf6ri.md
• raw/_consumed/2026-05-26/troponiniq_kb.md