Why AI coaching agents need an evidence packet before they advise.
A coaching agent should spend its context on the athlete conversation, not on re-parsing weeks of exports and guessing which number is safe to mention.
A good coach can look at messy athlete history and decide what matters. A language model can sound like it did the same thing. That is the problem.
If an AI coaching agent is drafting check-ins, program notes, or athlete replies, it needs more than raw workout files and a prompt that says "be careful." It needs a compact, deterministic answer from the data layer before it writes.
01The agent does not need all the data
More context is not the same as better grounding. Dumping a month of workouts, wellness logs, benchmarks, travel notes, soreness scores, and HRV readings into a prompt can make the agent slower and more expensive without making the advice more accurate.
The useful object is smaller. Before the agent writes, it should know the candidate driver, the outcome, the effect estimate, the confidence interval, the usable sample, the caveats, and the license on what it is allowed to say.
driver: "aerobic_decoupling"
outcome: "time_trial"
evidence: SUPPORTED (n=19, p<.001)
effect: +2.4% late-race pace
interval: [+0.78, +1.93]
caveat: "include fatigue flag"
license: ADVISE
That packet is short enough to fit beside the athlete conversation and strong enough to keep the agent from inventing its own statistical interpretation.
02The packet gives the model depth without giving it permission to bluff
Without a packet, the agent has to infer three different things at once: what the signal probably means, how strong the relationship is, and how cautiously to phrase the recommendation. That is too much freedom in a performance setting.
With a packet, the language model can do the part it is actually good at: translating a constrained finding into coach-quality language. It can write a check-in that sounds human while still carrying the interval, caveat, and uncertainty.
Deterministic code decides what the evidence allows. The agent decides how to explain it.
03Licenses matter more than labels
"Supported" is not enough. A coaching agent needs to know what action is allowed. AthDash attaches a license to each finding so the output can be capped before it reaches the athlete.
- ADVISE means the relationship is strong enough to use in a recommendation, with the interval and caveats included.
- HYPOTHESIZE means the pattern is useful enough to test, not strong enough to present as settled.
- WITHHOLD or DECLINE means the agent should not claim the relationship at all.
- BORROW means a cohort prior may guide observation, but it must not be dressed up as this athlete's own effect.
04This saves context as a side effect
The main value is better advice. The second value is lower context load. If the agent receives a pre-digested evidence packet, it does not need to spend tokens scanning every workout row, rebuilding a benchmark history, or rediscovering caveats that deterministic code already computed.
That matters when an agent is working across a roster. A small packet lets the model keep the athlete's goals, tone, recent conversation, and coach preference in context while still grounding the statistical claim.
05The coach stays in charge
The packet is not an autopilot. It is the reference point behind the coaching decision. The coach can accept it, ignore it, or use it as a prompt for the next N-of-1 test. The important part is that the agent does not get to invent the evidence while sounding confident.
That is the operating model AthDash is built for: coach decision first, evidence attached, agent wording constrained by the same packet.