AWS Builder Center

Builder Center

Second Eyes for Agents for Humans: an Professional Fashion Editorial Agent Built for the Brand Team that Knows How to Save Time and When Not to Guess

Second Eyes for Agents for Humans: an Professional Fashion Editorial Agent Built for the Brand Team that Knows How to Save Time and When Not to Guess

Built for the Agents for Humans Hackathon (Professional track) with the Strands Agents SDK. An marketing & communication editorial agent that help the brand team turn hundreds of creative assets into a human-ready shortlist.

Series: Agentic Workflow (1 article)

  1. 1
    Second Eyes for Agents for Humans: an Professional Fashion Editorial Agent Built for the Brand Team that Knows How to Save Time and When Not to Guess This article

The problem: 400 photos, four hours to deadline

A social editor gets handed a hard drive before a publishing deadline.
Nobody can look at four hundred photos, compare each one, and still make sound calls on portrait rights, product info, and channel fit. Rushing creates mistakes — and mistakes get expensive.

What we built: a shortlist, not a catalog

Second Eyes hands you a shortlist first. Given a publishing brief, it delivers 10–20 candidates with a reason and pending items per photo, plus coverage (which required scenes HAVE candidates, which are MISSING and stay missing — never guessed). The human picks the final three by instinct.
No hero ranking. No taste. No guessing.

How it works on AWS and Strands

  • Strands Agents SDK drives the loop: a Gemini-via-Vertex-AI provider model describes only; a pure deterministic route() function decides CLEAR / CONFLICT / UNKNOWN / NEW on frozen closed vocabularies (colour/item/shot, CJK substring scan). Structural honesty, not prompting.
  • Every model selection must carry an attached evidence_ref (frame / crop / tool-result ID) or it is discarded. Run records are validated JSON with token usage — including the thinking-token diagnostics that caught a real budget bug during the build.
  • The static Publishing Workroom UI (paper-toned editorial wall) presents SHORTLIST 18 / NEEDS REVIEW 7 / REMAINING 375 over a 404-file scanned corpus framed as a 400-photo scenario. Brief parsing, clustering, and coverage computation are honestly marked presentation-only in our architecture diagram.

The build story: multiple agent eyes

This project was built the way the product works. A builder agent scaffolded the repo, a reviewer agent attacked the routing contract (and found a real bug: disjoint fields wrongly clearing), a gate agent froze the vocabulary before any run, and a human owner made every judgment call the evidence couldn't support. No single agent — human or model — was allowed to assert what it wasn't given. The team that knows when not to guess, built the agent that knows when not to guess.
The demo video (3:25) shows the working shortlist, the no-brief gate, and the exact boundary where the machine stops and the human starts.
Repo: https://github.com/Crystal32378/second-eyes-agent
Live demo: https://crystal32378.github.io/second-eyes-agent/
# agents# advertising-marketing-industry# data-strategy# strands-agents# ai-assisted-workflows

Series: Agentic Workflow (1 article)

  1. 1
    Second Eyes for Agents for Humans: an Professional Fashion Editorial Agent Built for the Brand Team that Knows How to Save Time and When Not to Guess This article
Any opinions in this article are those of the individual author and may not reflect the opinions of AWS.
Enjoyed reading this content? Let the author know!

Your likes, comments, shares, and saves help creators reach more builders.

Loading recommendations

Loading article