What AI agents need from fitness metrics

If the user is an AI system, the bar changes.

Human-facing fitness products can get surprisingly far with proxy metrics. Time, score, rounds, placement, and streaks are all easy to understand. They compress a messy reality into something people can react to quickly.

Agents do not need that compression for emotional reasons.

They need a signal they can optimize.

Proxies are useful, but they are not the object

A finish time can tell you who won a workout.

It does not fully describe how the result was produced. It does not preserve pacing structure. It does not cleanly represent recovery, fade, or how output changed across segments.

For humans, that may be acceptable.

For software that needs to compare sessions, reason over changes, and make future decisions, it is not enough.

The system needs a measurement layer

This is the core WorkCapacity thesis:

before coaching systems get truly good, they need a measurement layer that represents the thing being optimized.

In this domain, that means work over time.

That is why Power Oracle matters. It turns workout structure, athlete inputs, and elapsed time into computable outputs such as mechanical work and average power. More importantly, it preserves enough structure that downstream systems can reason from the result instead of flattening everything into a final score.

What agents actually want

An optimization system generally wants a few properties from a metric:

• objective definition
• repeatable computation
• enough structure to support comparison over time
• machine-readable output
• clear constraints and assumptions

If one of those is missing, the system starts guessing.

That is usually where low-trust recommendations come from. The issue is not always the model. Sometimes the issue is that the target being optimized was too lossy from the start.

Why this matters commercially

A lot of software is built as if the end user will always be a person sitting in a dashboard.

That assumption is weakening.

As more workflows become agent-assisted, products that expose direct and reliable signals gain leverage. They can be integrated into planning systems, evaluation loops, and decision engines without first being translated out of a human-only UX layer.

That is one reason WorkCapacity starts with measurement instead of coaching copy or community mechanics.

The signal has to exist before the higher-level system becomes trustworthy.

The practical takeaway

If you want agents to improve fitness outcomes, do not start by asking what interface they need.

Start by asking what they can measure, compare, and optimize without ambiguity.

That is the layer WorkCapacity is building.