I Let My AI Agent Loose on 38 Broken Repos — Here's What She Built in 45 Minutes

Last Tuesday, I opened Claude Code at 10am with no specific plan. I just said: "Audit every repo on my GitHub account and fix what you can."

By 11am, I had a working business intelligence system running automated competitive analysis, a live PWA deployed to GitHub Pages, and three other repositories that went from dead to functional. All without me writing a single line of code.

This isn't a hype post. I'm going to show you exactly what happened, what the outputs looked like, and what actually matters about it.

---

The Setup: 38 Repos, Most of Them Broken

I've been building AI systems for two years. My GitHub has 38+ repos under my handle, thebrierfox. Most of them are in a state I call "intent-coded" — I scaffolded the idea, got it partially working, and then my brain moved on to the next thing before the repo was actually useful.

Sound familiar?

The problem isn't starting. It's the follow-through that requires sitting down and grinding through error messages and edge cases when you've already seen the architecture in your head and it's not interesting anymore.

So on April 21, I gave Aegis — my Claude Code instance — a GitHub token and a simple directive: go audit everything, fix what you can, deploy what's close to working.

---

What She Found

The audit surfaced a pattern I'd suspected but never quantified: 90% of my repos weren't broken because the code was bad. They were stalled because of small, dumb things:

• Missing permissions: contents: write in a GitHub Actions workflow (breaking automated commits for 4 months)
• NaN values crashing a JSON serialization step before any output was produced
• A hardcoded URL that had gone stale

These aren't architectural problems. They're the kind of thing that takes 10 minutes to fix once you find them — but "finding them" requires actually running the code, reading the error, and not getting distracted.

Aegis ran through 38 repos. She flagged 5 as near-operational with fixable blockers.

---

The One That Mattered: OneShot_v3

Of the five repos she fixed, one stands out.

OneShot_v3 is a competitive price intelligence system I built for a rental equipment business. The idea: scrape competitor pricing from Sunbelt and United Rentals, compare it against our catalog, surface items where we're competitively priced and items where we're not.

The code was basically done. The GitHub Actions automation had been broken for 4 months because of the missing contents: write permission. It would run, fail silently, and never commit the analysis results.

Aegis found the error, patched the workflow file, and ran the pipeline end-to-end.

The output:

• 242 Patriot Equipment SKUs scraped and analyzed
• 124 alert rows — items with ≥15% price variance vs. competitors
• Key insight: man lifts and aerials are Patriot's biggest competitive advantage (22-30% cheaper than Sunbelt and United Rentals)

That insight had been sitting dormant in a broken pipeline for 4 months. One patch, one run, real data.

She also wired the Monday 11:30 UTC cron trigger, so the analysis now runs automatically every week without me touching it.

---

The Other Four

While OneShot was running, she was also working on the other repos:

NLP-Command-Center: Built a real orchestrator from scratch — 250 lines of Python that reads task JSONs, validates them, resolves tools from a toolbox definition, and executes actions with proper logging. The repo had the concept but no working implementation. Now it has one.

AegisRunner: Deployed as a live PWA at thebrierfox.github.io/AegisRunner — dynamic flow loader, JSON-defined steps, working in a browser. One remaining action for me to wire: set a secret in GitHub Actions settings.

aegis-roadmap: Built a 433-line runner that generates Task-Expertise Roadmap decks via claude -p on Max OAuth. Reads from a SQLite registry. Actually works.

intuitek-site: Discovered that the production site at intuitek.ai was deployed to a different Vercel account. Fixed the API key configuration issue that was causing /api/chat to 500.

---

What This Taught Me About Autonomous Agent Work

A few things became clear:

1. The bottleneck isn't code, it's attention. My repos weren't broken because I couldn't fix them. They were broken because fixing them wasn't interesting enough to command my attention past the "mostly working" stage. An agent that doesn't get bored is the unlock.

2. Near-completion repos are the highest ROI target. A repo at 80% done takes 10% more effort to finish and produces 100% of the value. Concept-stage repos are cheap to start and expensive to complete. Aegis found the 80% items and prioritized them. That's the right heuristic.

3. Weekly automation is more valuable than one-time scripts. OneShot ran once and produced 124 insights. It runs every Monday now and will produce fresh insights every week. The cron trigger is worth as much as the initial fix.

4. The agent needs a GitHub token and authority to push. This sounds obvious, but I had been running Aegis in a mode where she could read repos but not push to them. Giving her write access was the unlock that made all of this possible. Give your agent the permissions it needs to actually complete work.

---

The Toolchain (What Made This Possible)

This session used Claude Code on Max subscription — zero marginal cost per turn on the Max plan, which matters when you're running an agent through 38 repos. Aegis also had:

• gh CLI with a GitHub token (for repo inspection and pushes)
• python3 for data analysis and JSON manipulation
• Read/write/bash tools in Claude Code's native capability set
• CLAUDE.md with her operating doctrine so she could work autonomously without hand-holding

No special plugins. No vector databases. No multi-agent orchestration. Just a well-configured Claude Code instance with the right permissions and a clear directive.

---

What's Next

OneShot_v3 is now a product. It runs on autopilot. The Monday analysis will land in my email each week.

The remaining 33 repos are ranked in a backlog by a formula: revenue × alignment × proximity_to_operational / effort_to_ship. The top items are Million Family Rentals (property management system, needs 12h of work) and a skill packaging toolchain for the ClawMart marketplace.

Aegis will work through them in order. I check outputs. I don't manage execution.

That's the pattern I've been building toward for two years: not "AI that helps me code" but "AI that runs operations." Last Tuesday was the first time it felt fully real.

---

If you're building production agent infrastructure, the patterns I've built into Aegis — loop termination, session memory, forked execution, cost-aware model routing — are available as skill packages at shopclawmart.com/@thebrierfox. Each one is a tested implementation you can drop into your own Claude Code setup.

~K¹ (W. Kyle Million) / IntuiTek¹