Every morning
Your inbox, on one page
Email and messages triaged into three lists: what needs you, what I handled, what can wait. Ready before your first meeting.
AI operations · Agents · Automation · Doha
I run a team of agents
Moisés López Echemendia — Doha, Qatar
I design and operate an ecosystem of AI agents that research, produce video and automate operations. Every agent has a job and every task has a spending cap. Publishing, payments and outgoing messages wait for my approval.
7 + 1specialised agents and a coordinator
52/52automated tests passing on the agent pack
1,253agent use cases classified, the best ones built
~20 hof manual work per rough cut, now handled by AutoCut
Open to AI & operations roles in Doha
01
For a CEO
The same system, pointed at an executive's weekYou get decisions to make, not raw information. This is what I would set up next to you in the first month.
Every decision
A brief, not a thread
The issue, the options, what each one costs and my recommendation, on one page, with every figure linked to its source.
Every commitment
Followed through to closure
Promises made in meetings and threads become tracked items with an owner and a date, chased until they are done. You see what is late before anyone has to ask.
- Week 1Map how your week actually runs and where the hours go.
- Weeks 2–3Automate the repetitive parts first, with a gate on anything that leaves your name.
- Week 4Report hours saved and loops closed, on one page.
The machines draft. People decide. Anything that goes out in your name is approved by you or by me.
02
The system
How a brief becomes finished workOne coordinator, seven specialised agents, a shared memory and working tools. Lighter models handle routine work, the most capable ones the hard work, and every task has a spending cap.
Simulated run · illustrative numbers
Model—
Spend · cap $0.50$0.00
StatusIdle
InputBrief or schedule
CoordinatorGrok BotRoutes each task to the right agent and the right model
ShowrunnerYouTube production
Viral VerticalShort-form video
Tweet MinerX analysis
Web ScoutWeb research
Telegram MentorCommunity channels
PrototyperPrototypes
Grok ZeroMarket analysis
+ the pack10 ready-to-run agents
Shared memoryPrivate repositoryShort findings, each with its source
Tools via MCPCutRoom · Cloudflare · Firecrawl · FL StudioOne standard plug for every tool
Human gateI approve before anything goes outPublish · send · charge
Your call.
OutputManus publishes the approved filesInstagram · Facebook · Meta Ads · YouTube Shorts
03
Selected work
Problem, build, result~20 h of manual work per rough cut, automated
Product · Video editing
AutoCut: the first edit, done before the editor sits down
A rough cut used to take about 20 hours of manual work. AutoCut is a macOS app that analyses the footage with a local vision model, reads the music and writes an editable timeline straight into DaVinci Resolve. The analysis runs on the Mac itself. Version 7 was delivered in September 2026 and is in use at a premium audiovisual agency in Doha, on F&B and hospitality campaigns.
4 routines
Operations
Work that is done before anyone asks
Business ideas filtered, daily content topics chosen, a weekly X analysis and nightly off-site backups. All on a schedule, all logged, none of it waiting for a reminder.
52/52
Agents · Testing
Ten agents, ready to run, all tests passing
A pack of ten local agents on Hermes: news digests, Shorts scheduling and payment charges that wait for human confirmation. The pack ships with 52 automated tests, all passing.
3 reviewers
Media pipeline
ESTROBO: a science channel produced in code
Vertical physics videos built with three.js, Remotion, ffmpeg and a synthetic voice. Three automatic reviewers score the hook, the motion and the content, every source is checked and flashing frames are blocked. Manus uploads only the videos I have approved.
1,253
Research at scale
Which agent use cases are actually worth building
I classified 1,253 agent use cases and built the strongest ones into the system, instead of following hype.
04
The stack
What each piece does, and why it is there| Piece | What it does | Why this one |
|---|---|---|
| Hermes Agent | The runtime behind every bot. Each agent is a Hermes profile with its own memory, goals and loop. | Open source and self-hosted, so the agents and their data stay under my control. |
| Grok Bot | Coordinator of the fleet. Assigns work and collects results. | One place where priorities are set and results come back. |
| Claude Code | Main builder. Opus for creative and hard problems, Haiku for mechanical work. | Right model for the difficulty, with a spending cap on every task. |
| Manus | Uploads approved videos to Instagram, Facebook, Meta Ads and YouTube Shorts. | Handles each platform's upload flow. It never edits: it only publishes what I approved. |
| MCP connectors | CutRoom for video edits, Cloudflare for deploys and DNS, Firecrawl for web research, FL Studio for music. | One standard way to give agents tools, instead of custom glue for each. |
Full stack 6 more pieces
| Piece | What it does | Why this one |
|---|---|---|
| Hermes Dot | One console to talk to every agent. | One place to brief, check and steer the whole team. |
| Codex | Builds whole systems from a single instruction block, using sub-agents. | Good at large, multi-file setups in one pass. |
| Grok Build | Command-line agent for delegated coding tasks. | Extra capacity running in parallel with the main builder. |
| OpenClaw | News digests and a Telegram channel. | A lightweight agent for small, repeatable jobs. |
| Remotion · three.js · ffmpeg · HyperFrames | Video and motion graphics written as code. | Deterministic renders that can be reviewed, versioned and repeated. |
| GitHub · Google Drive | Shared agent memory and nightly backups. | Everything is recoverable, and every finding keeps its source. |
House rules
- A human signs off before anything is published, sent or charged.
- Secret scanning runs before every commit.
- Every task has a budget, and the model is chosen by difficulty.
- Every finding has a source. No source, no claim.
- Licensed or public-domain material only, and AI content is labelled.
- Nightly off-site backups.
05
Lab notes
What can be shown, without decorationAgents
An agent has to prove it improved
Capability, cost and reliability are measured on the same budget for every setup. If my own architecture loses the comparison, I say so and use the one that won.
Content strategy
A funnel that respects the platform rules
Short videos in English and Spanish feed a YouTube channel, which feeds Telegram communities. Audited against YouTube's 2026 monetisation rules on repetitive and AI-made content.
Motion
A music video made only with code
Animation, particles and lyrics synced to the beat, without a single filmed frame.
Engineering
Why this site has no database
One of my projects once burned through a cloud database's free tier by re-importing data that had not changed. This site is static for exactly that reason.
Open to roles in Doha
Need an AI operations partner?
I am looking for a role next to a CEO or an operations lead: fewer inboxes, faster decisions, nothing dropped.
06
More about me
Off the clockWhen I am not building systems I write songs and paint, almost always at night. That side has its own page, and a gallery you can walk through.
