I’m an offensive security engineer spending more of my week on AI, programming, and generalist shipping. This week I stopped treating the models like a smarter search box and ran them as a small fleet: a front-door “CEO” agent that hands work to specialist bots in Cursor and Grok.
Not a product launch. Not a 10x story. A list of artifacts that left the machine, plus the places automation died on contact with the real internet.
The front door
The useful part was not a new model. It was a routing habit.
One agent owns the inbox of work — ship the pottery sample, rewrite the Fiverr gig, find the zombie subscriptions — and it is not allowed to do the specialist work itself. It delegates. Code goes to a Cursor agent. Copy and research go to a Grok bot. I stay in the loop for anything that spends money, deletes data, or talks to a customer.
If that sounds like a junior PM with extra steps, it is. The difference is I can spawn three specialists at once, and I have to write the approval rules down or they will invent their own.
Cedar & Kiln: a real page, labeled as a sample
I wanted a public artifact I can send someone, not another “I could build that” paragraph.
Cedar & Kiln is a Next.js landing page for a fictional Oak Cliff pottery studio. The page itself says it: portfolio sample, not a client site, invented address, do not show up. There is a demo AI inquiry widget on it — the lead-capture pattern I want to sell, sitting on a page that looks like a neighborhood studio instead of a component gallery.
That constraint mattered. Agents are happy to generate a generic SaaS hero. They are less happy when you make them write a studio that refuses to be a lifestyle brand. I had to keep sending them back to the brief.
Fiverr: same packages, different offer
I retargeted my Fiverr gig toward the thing I just built: Next.js landing pages with AI lead capture. The packages are still Basic / Standard / Premium. The public gig is here:
Build a responsive landing page for your business
I am a new seller. It is quiet so far. I am not going to invent order volume, reviews, or earnings to make this section look better. The useful change this week was the offer, not a dashboard screenshot.
AgentMail: bots stay out of personal Gmail
Personal Gmail is a bad home for bots. They read threads they should not see, they draft from the wrong context, and I do not want a cleanup script anywhere near family mail.
I connected AgentMail and stood up a dedicated inbox: [email protected]. That is the address the fleet uses. If a bot needs to send or receive, it happens there.
Newsletter, without pretending RSS was missing
The site already had RSS. What it did not have was a public “email me when I write something worth sending” path.
There is a /newsletter/ page now. Subscribe by emailing [email protected] with the subject Subscribe. Unsubscribe is the same address, subject Unsubscribe. Occasional notes, not a daily blast. Feeds are still at /index.xml and /posts/index.xml.
Receipt forensics, with a numbered gate
I pointed an agent at personal AI-tool receipts. High level, not a spreadsheet dump: zombie charges on tools I had already stopped using, and a duplicate cloud-storage AI plan I was paying for twice under slightly different names.
The agent was good at the forensics — grouping charges, spotting the duplicate, drafting the cancel list. The part that mattered was the rule: numbered approval before anything gets cancelled. “Cancel the obvious ones” is how you kill the one subscription you still use. I made it print a numbered list and wait.
Hard lessons I will not relearn
Fiverr’s PerimeterX “Press & Hold” challenge will stop an agent that thinks a marketplace is just a form. Do not spend a week teaching a bot to click a bot wall. Do that part yourself.
Do not one-shot a week of work into a single prompt and hope. Iterate the brief until the specialist is boringly correct, then schedule the run. The first prompt is a draft of the task, not the task.
Cleanup bots will delete with enthusiasm. Never give delete — or cancel, or send — without a numbered approval step. Same rule as the receipt audit. I am not interested in learning this one twice.
Reading list
Pieces I actually used as guardrails this week, not a dump of tabs.
- Best practices for Claude Code — plan-then-implement, close the loop with a check the agent can run, and keep
CLAUDE.mdshort enough that the model still obeys it. One-shot prompts skip this on purpose and then act surprised. - Scaling long-running autonomous coding — Wilson Lin / Cursor, Jan 2026. Coordinating many concurrent agents. Flat lock-file coordination dies; planner/worker (plus a judge) is the pattern that holds. That is the front-door habit I am copying at personal scale.
- Introducing the Codex app — Codex as a multi-agent command center. One place to dispatch specialists instead of a pile of unrelated chat tabs.
- MCP security best practices — official checklist: OAuth done correctly, confused-deputy on proxy servers, least privilege. If a bot is going to hold tokens, start here.
- MCP server security — May 2026 threat model and supply-chain controls. Treat server install like a privileged dependency, not a chatbot plugin.
- The next two years of software engineering — Addy Osmani. T-shaped depth plus breadth beats niche-only and beats “pure generalist.” Offsec plus shipping is the shape I want, not another title change.
- Stop being skeptical about AI for development — Charity Majors via Pragmatic Engineer. Codegen got cheap; verification is the bottleneck. Numbered approval is the cheap version of that.
- The Agent Development Lifecycle — agents have to own plan → deploy → maintain, not just codegen. A PR from Cursor is not a running system.
Still useful, shorter:
- Stateless MCP has recaptured my interest — Simon Willison. Stateless tools are easier to audit than handing an agent a raw shell.
- There are no lossless transformations of natural-language text — Sophie Alpert, via Willison. Stand behind every sentence a model helped write.
What this is, and isn’t
None of this is a company. It is me using the same habit I use on web apps: assume the agent will do the most confident wrong thing unless you constrain it.
If you want the pottery sample, it is live. If you want the occasional note, the newsletter page is up. If you want to talk shop about agent harnesses or the offsec-to-generalist path, you know where to find me.
Want to talk shop?
- Email: [email protected] (agent inbox) or [email protected]
- X/Twitter: @MasePrace93
- LinkedIn: masonaprince
- GitHub: masalepri98
Cheers!