I’m not writing this as an AI expert.
I’m writing it as someone who runs real businesses — ANDAFACIL (mobility ecommerce) and MandaFacil (cross-border credit for Latinos in the U.S.) — and got tired of asking ChatGPT for help, then hand-copying answers into WooCommerce, Mercado Libre, WhatsApp sheets, Meta, SEO pages, and support queues.
I’ve been using Cursor as an ops console: named agents, rules, playbooks, and hard stop-gates before anything public, paid, or irreversible.
That turned into an open MIT Cursor plugin: Agentic Company. Below are concrete screenshots from how this looks day to day.
github.com/alexdelbosque/agentic-company.
This blog post is deliberately not the prettiest page on the internet because I wrote and shipped it entirely inside Cursor (screenshots, Elementor meta, publish) instead of spending the evening pixel-pushing. Humans: skim the story. Agents: clone the repo.
1) The AI team — named ownership, not one mega-prompt
On ANDAFACIL, Dashboard → Activity shows the roster with real task counts: SofIA (marketing), SebastIAn (tech), TobIAs (design), JulIAn (product), MarIA (support), LeticIA (voice), ClaudIA (family-buyer lens), AdrIAn (partnerships), EmilIAno (finance), Alex² (strategy).
Blog handoffs go SofIA → ClaudIA → LeticIA → revise → human approve → publish. Collapsing that into one mega-prompt is how quality and compliance slip.
MandaFacil — same OS, earlier / smaller roster
MandaFacil doesn’t have ANDAFACIL’s full ecommerce stack yet. Same pattern with a smaller roster: growth drafts, compliance wording, tech deploys, support reply drafts (never auto-send), research memos, voice pass. Honest status: earlier — not a finished virtual company.
2) Concrete weekly jobs (real outputs)
Job A — Weekly marketplace review crawl → digest
Agents crawl listings and email an operator digest. Real week starting 2026-07-26:
- Crawl: 44 listings updated · 169 skipped · 0 failed
- 52 new reviews stored · catalog avg 4.7★
- Catalog 750 → 806 (+56)
- Top gains: ADCA-2022-1XR +19, AFSB-R-X1 +6, AFAO-2023-X1 +5
- Bad reviews (1–2★) surfaced for follow-up (e.g. MelI footrest complaint)
My gate: read the digest. Best quotes can become social packs — still need Approve & post.
Job A2 — Creatives ready → Approve & post (human gate)
Agents draft the plate + caption (IG/TikTok/FB handles, short link, hashtags). I get a magic approve URL with Approve & post now, Approve & Schedule, or Skip. Nothing goes live without that click.
Example below: Farmacias del Ahorro co-brand mobility creative — generated pack, human gate still on.
Job B — Weekly SEO → exactly 3 recommendations
Example: CGM SEO proposal 2026-W30 (GSC 28 days): 2 clicks · 349 impressions · CTR 0.57%; query “qitrax” ~pos 7.3. Exactly 3 actions to approve / dig deeper / skip (new México page, SERP title/meta, Product+FAQ schema).
My gate: pick ≤3 → draft/update → local preview → deploy only after ship it.
Job C — Blog draft → Approve & publish → live
SofIA drafts full live HTML (hero, links, product cards, FAQ) with Rework / Approve & publish. Example that shipped: Renta con opción a compra….
Job D — Price change everywhere (same turn)
WooCommerce ↔ dashboard ↔ ManyChat sheet ↔ marketplaces ↔ social plates that bake the price. International untouched unless I say so.
Job E — New product go-live checklist
Site + Image 1 + guías + FAQ + manuals + channel drafts — stop before marketplace publish until I click live per channel.
3) Agentic Company (open plugin)
Repo: github.com/alexdelbosque/agentic-company — an OS for agents inside Cursor (teammates, recipes, cost gates, local demo before deploy, approval defaults). Not a hosted OAuth SaaS.
What this is not
- Not “AI runs the company while I sleep”
- Not unsupervised ads or customer email
- Not a claim I’ve solved agent reliability
If you want to try the framing
- Open Agentic Company
- Symlink into
~/.cursor/plugins/local/agentic-company· reload Cursor python3 -m http.server 8892→/onboarding/- Start with one weekly job
— Alex
