Most AI agents are a for-loop in a trench coat

SeptHQ is seven AI agents that run the weekly grind of social marketing for a brand: research, strategy, copy, design, quality control, distribution, measurement. When I first wired them together, five of the seven did not actually think. They looked like they did.

What each one really did was build a block of text and hand it to the same content generator, which returns slides. A slide array cannot carry a relevance score or a quality verdict, so the code simply invented those numbers. The QC agent approved everything with a hardcoded seven out of ten. The research agent ranked trends by their position in an array. The analyst returned a random number dressed up as engagement.

It demoed beautifully. That is the trap. A row of agents reporting confident numbers reads as intelligence, and none of it was real. A brand would have made decisions on it.

The fix was structural, not cosmetic. Each thinking agent got its own structured-output call: a raw model call that must return JSON matching a schema, and retries with the validation error fed back until it complies. When the model returns the score, the score is real. When it returns ranked opportunities with reasons, the ranking is real. The fake constants are gone because there is nowhere left for them to hide.

The quality agent now actually judges a draft against the brand, scores each dimension, weights brand fit double, and gates: anything that fails routes back through one automatic refinement before a person ever sees it. Nothing posts without a human saying so, because a strategist's judgment is the thing an agency is actually selling.

The honest part: the planning half is real and runnable end to end, from live signal to an approved week of drafts. The autopilot half, actually posting to a channel and reading the results back to learn from them, is not built. I shipped the part that works and I am not going to describe the machine as finished when half of it is a diagram.