Showcase¶

Bricolage agents in action — real sessions, edited for brevity.

1. Master → Planner → Guide → Assembler: a blog in one session¶

User: I want to build a blog with AI summarisation. Free only.

Master:

Good goal. Let me check what we've got. Compute domain is blessed and verified 2026-03-23. I'll hand you to the planner.

Planner output:

domain: compute
goal: AI-summarising blog
pattern: blog-ai
resources:
  - name: GitHub
    role: source repo + Actions CI
    has_guide: true
  - name: Vercel
    role: hosting + edge functions
    has_guide: true
  - name: Supabase
    role: posts + summaries table
    has_guide: true
  - name: OpenAI
    role: gpt-4o-mini for summarisation
    has_guide: true
steps:
  - create public GitHub repo
  - scaffold Next.js with Supabase + OpenAI SDK
  - wire Vercel deploy
  - write the /api/summarise edge function
  - verify end-to-end with a seed post
warnings:
  - OpenAI free credit is classified `trial` — will decay once the $5 starter is used
  - Supabase free tier: 500 MB, pauses after 7 days inactive
estimated_effort: 45 min

Guide: Walked through GitHub signup → Vercel OAuth → Supabase project → OpenAI API key. Stored four tokens via ubc_store_access with regex validation. None of the four were ever printed in the chat.

Assembler: Scaffolded create-next-app, wrote lib/supabase.ts and app/api/summarise/route.ts, ran vercel --prod, verified /, /posts, and /posts/seed-post all returned 200.

Total time: 38 minutes from "I want a blog" to https://brunos-ai-blog.vercel.app/ being live.

2. Discovery: scaffolding a domain on demand¶

User: I want to learn data science for free but I don't know where to start.

Master: No matching domain. Delegating to discovery.

Discovery: 1. Called ubc_create_domain with id: education, trust_level: user_scaffolded. 2. Ran 14 web searches. Verified free tiers on: fast.ai, MIT OCW, Kaggle Learn, freeCodeCamp, Coursera audit track, StatQuest, 3Blue1Brown, The Odin Project, Full Stack Open, Harvard CS50, DeepLearning.AI (audit), Google ML Crash Course. 3. Wrote 12 resource YAMLs under domains/education/resources/ and a bulk entry for 34 adjacent resources in catalog.yaml. 4. Authored the starter pattern ds-foundations.yaml: Khan Academy math → StatQuest stats → Kaggle Learn Python → fast.ai → Kaggle competition. 5. Reported back: "Domain education scaffolded. trust_level: user_scaffolded — the catalog is my best research, not a maintainer's review."

That honesty flag is the feature.

3. Infra: PicoClaw on GitHub Actions¶

User: I want the catalog to re-verify itself weekly.

Infra: Recommended PicoClaw on GitHub Actions because the task is periodic and the user's whole project is a public repo.

Wrote .github/workflows/catalog-refresh.yaml:

name: Catalog Refresh
on:
  schedule:
    - cron: "0 14 * * 1"  # Mondays 14:00 UTC
jobs:
  refresh:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - uses: actions/setup-node@v4
        with:
          node-version: 20
      - run: npm ci
      - run: npx tsx tools/scripts/verify-catalog.ts --write
      - uses: peter-evans/create-pull-request@v6
        with:
          commit-message: "chore: refresh catalog staleness tags"
          branch: auto/catalog-refresh
          title: "Weekly catalog refresh"

First run took 4 minutes of the monthly 2,000-minute budget. PR arrived with 11 verified_at bumps and two newly-stale entries auto-flagged.

Screenshots¶

The deployed site, the encrypted-access directory listing, and the pre-push grep check all live in /04-Showcase/images/ in the repo. (Placeholder — add screenshots before submitting to the cohort.)

What this is proof of¶

Six agents, one protocol.
411 resources, every one verified and dated.
A credential store that encrypts everything and never prints a secret.
A domain model that lets the system grow on demand without pretending the new parts are blessed.

This is the fortress. The site you're reading is the window into it.