Portfolio Agentic Loop System

An adversarial agentic system that generates improved portfolio variations through iterative feedback loops between improvement and hiring manager agents.

What It Does

The system creates portfolio variations optimized for senior design and product roles by:

1. Improvement Agent - Analyzes the baseline portfolio and proposes enhancements
2. Hiring Manager Agents - Evaluate each variation from the perspective of 10 target roles
3. Feedback Loop - Iterations continue until 10 approved variations are generated

Target Roles

See target-roles.txt for the complete list. Currently optimized for:

• Head of Product Design
• Senior UX Researcher
• Product Manager (Design-Led)
• Director of Design
• Lead Product Designer
• VP Product Design
• Design Systems Lead
• Staff Product Designer
• Principal Designer
• Head of User Research

Files

Core System

• coordinator.mjs - Main orchestrator that runs the agentic loop
• target-roles.txt - List of roles to optimize for
• loop-state.json - Current state (resumable across runs)

Agent Prompts

• agents/improvement-agent-prompt.md - Template for improvement agent
• agents/hiring-manager-prompt.md - Template for evaluation agent

Legacy/Alternative Implementations

• agentic-loop.js - Node.js implementation with agent framework
• run-agentic-loop.sh - Shell script orchestrator

Output Structure

Each portfolio variation is saved in archive/portfolio-agentic-{timestamp}/:

• manifest.json - Metadata, improvements, and evaluations
• index.njk - Eleventy page with variation content

Summary Files

• archive/agentic-loop-summary.json - Complete run summary
• archive/index.njk - Archive landing page

Running the System

Full Run

cd /workspace/portfolio
node coordinator.mjs

This will:

• Generate 10 portfolio variations (with adversarial feedback)
• Create pages in archive/portfolio-agentic-*/
• Update the archive index
• Save run summary

Resume from State

If interrupted, re-run node coordinator.mjs - it will resume from loop-state.json.

Fresh Start

rm loop-state.json
node coordinator.mjs

How It Works

Iteration Loop

1. Generate Variation - Improvement agent creates new portfolio version
2. Evaluate - 10 hiring manager agents score it (0-100) for their role
3. Decision - Average score ≥75 = approved, <75 = iterate with feedback
4. Repeat - Until 10 approved variations exist

Approval Criteria

• Each hiring manager scores 0-100
• Average must be ≥75 to approve
• Rejected variations feed back into next iteration
• Maximum 2 rejections observed before approval in current run

Results

Last run completed: 2026-09-07

• Total iterations: 12
• Variations generated: 10
• Rejection rate: 16.7% (2 rejected, 10 approved)
• Average score range: 76.1 - 82.3 / 100
• Highest scoring: Variation 10 (82.3/100, focus: Head of Product Design)

View all variations: /portfolio/archive/

Baseline

The system improves upon the baseline portfolio at: /workspace/portfolio/archive/portfolio-2026-09-02T151008/

Eleventy Integration

Generated variations are:

• Excluded from collections (eleventyExcludeFromCollections: true)
• Set to noindex, nofollow
• Available at /portfolio/archive/{variation-id}/

Build normally with Eleventy:

npx @11ty/eleventy --serve

Future Enhancements

Current implementation uses simulated agents. Next steps:

1. Real Task Agents - Replace simulated evaluations with actual Cloud Agent Task calls
2. Content Generation - Have agents write actual enhanced project descriptions
3. Deeper Feedback Loop - Feed specific hiring manager concerns back to improvement prompts
4. Role-Specific Variations - Generate targeted versions per role instead of optimizing for all
5. A/B Testing - Track which variations perform better in real hiring scenarios