Record it, watch it back, rebuild: the loop behind a pitch and an AI agent

In late August 2026, Structed joined cohort 4 of the Grishin Robotics accelerator in London: 22 teams, 30 days, and a demo day on 17 September with three minutes each on stage.

Marina, Structed's founder, said the most useful part of the program wasn't the advice. It was a loop: pitch, get recorded, get feedback, watch yourself back. Rewatching, she could see exactly where the story broke: where she'd piled on details, and where something perfectly clear in her head wasn't clear to anyone else.

The same loop, for an agent

An AI agent has the same blind spot as a founder rehearsing alone. It sounds fine from the inside. You only see where it breaks when you watch a full run from the outside.

So the agent gets its own version of the loop:

• Pitch: run a whole interview with a test client, a persona with known facts, written from scratch or copied from a real conversation, mistakes included.
• Record: keep the full trace, every question, every answer, every decision.
• Feedback: score it against what the persona knew. Which facts made it into the requirements, which questions were wasted.
• Watch it back: read the transcript like a stranger would, and fix what broke.

Then run it again, and compare run against run.

Why it beats tuning the prompt

Tuning a prompt against one example feels productive and rarely is. A loop over whole runs shows you the problems that only appear in sequence: a question asked twice, a fact given early and lost late, an answer that's fine alone but wrong in context.

That's also why the loop has to be cheap to repeat. If running it takes a day, you'll do it before demo day and never again.