Zero to Agents

My name is Liberato Aguilar, founder of Aguilabs. Welcome back to Aguilabs Signal.

Big announcement: I'm making a free AI course.

It has been about five weeks, since issue #03. Two stories I promised for this one, Meta's ad blitz and the AI-server corridor, got bumped as the story keeps developing. They are still coming.

I'm also trying something new for this issue, I will keep the main sections on the email version of the newsletter and link to more detailed stories on my website. Feedback appreciated.

Let's get started.

Jump to: Zero to Agents · New models · Personal agents · America.gov · The blimp · Project Jupiter · The trading robot

I use agents every day, and you keep hearing about them: this issue alone mentions three companies offering you a personal agent. So I'm making the from-zero explanation. Zero to Agents is a free video course in English and Spanish, animated and narrated by me, made for everyday people. No tech knowledge needed!

The promise is: "Go from your first AI conversation to practical agents that can help with real work and everyday tasks."

The plan is 18 episodes in five parts:

• The model. What a text model is and how to ask it for something useful.
• More than text. Models that see, speak, draw and search.
• Giving it tools. The model stops chatting about the work and starts doing it.
• Many agents. One agent becomes an organized team.
• Autonomy. From a request to building, testing, and shipping automatically.

Every episode works on one small app, a meal-prep tracker, and each hands more of the building to the machine. It starts with copying code out of a chat window. It ends with someone asking for a feature from their phone while the software builds, tests and ships it, with a person approving changes along the way.

Where it stands: in production. Episodes 1 and 2 are done, and coming this week. The episode list may change. Episodes will be announced to newsletter subscribers via email, and the course page is aguilabs.com/zero-to-agents.

Why you care: agents are arriving whether or not anybody explains them. I'd rather you know the capabilities and risks that come with autonomous AI agents.

Nine releases worth knowing about since the last issue, and four of them list at $2 in, $10 out. The newest is Anthropic's Claude Haiku 5.5, out October 7 from $0.10 in, $0.50 out. That fills in the 5.5 line with Opus, Sonnet and Haiku. Only Fable 5.5 is missing: the top model is still Fable 5.1, and Anthropic has not announced a 5.5.

Opus 5.5 came in at $4 / $20, down from $5 / $25. OpenAI's and xAI's prices come from trade press, not their own pages. Argon is going first to what Google calls "trusted cyber defenders." Haiku 5.5's price is for prompts up to 100,000 tokens ($0.50 / $2.50 above that), down from $1 / $5 for Haiku 4.5.

The time between releases is shrinking fast.

Credits to @jfonsecarivera on X who posted this chart on September 29. Each point is a new model from Anthropic or OpenAI, placed by the days since the previous release from either lab. We got a new model every 70 days or so in early 2024, that is down to every 11 days now.

A separate count of five US labs points the same direction: ARK Invest's analysis of Artificial Analysis data, reported in April, puts the median gap at 37.5 days in 2023, 13.5 in 2024, 17 in 2025 and 11 in early 2026.

So just how much intelligence does AI contribute?

This one is a back-of-the-envelope estimate where I show cumulative IQ of all humanity vs AI IQ.

Blue is humanity: every living person times average IQ. Red is AI: the world's daily tokens (chunks of text) divided by what one human mind gets through in a day, all weighted at IQ 140. Its headline for September 2026: AI is about 23% of humanity's sum, with a band of 7% to 68%. At this rate AI will overtake the sum of human intelligence by 2028.

We are seeing it in math. If you saw headlines that AI solved a million-dollar math problem: on September 8 OpenAI released a computer-checked proof for one version of the Navier-Stokes problem, a famous open question about how fluids move. The Clay Mathematics Institute, which offers the prize, says the problem "has apparently been settled" and is under review.

On alignment. The rise in AI capabilities and intelligence does not mean you should be afraid. AI is not going to take over the world anytime soon. It is a tool that will help your business succeed if you know how to use it.

Why you care:

• Buy month to month. A tool you picked in spring is a few releases old by fall.
• Smarter models = business efficiency gains. Learn how to leverage AI capabilities to optimize and grow your business.

A chatbot answers. An agent does work. It gets its own cloud computer, its own browser and your logins, and it keeps working after you close the laptop. We are pretty close to everyone having a personal Jarvis now.

Muse is Meta's, announced September 8. It can "open a browser, fill out forms, and negotiate on their behalf." Muse will get its own email address and named shopping partners including Shopify, Stripe, PayPal and Walmart. Meta calls it "free for most of what people need." Be careful, Meta offers this free because all your interactions train Meta's models unless you opt out.

Dots are OpenAI's response. Announced September 29: always-on agents running on GPT-6 Astra, with connections to more than 4,000 apps, reached through ChatGPT, Slack or Teams. Only those on ChatGPT Pro (from $100 a month) and Business Premium get access for now. You can set read-only access and approval rules.

Bot is Grok Bot, from xAI, in beta since August 11. Named agents sign into company software and coordinate in a group chat. It is built more for teams than a one-person shop. The catch: all of your bots share one computer, cookies and credentials included, so a mistake by one bot may affect others.

Competition is rising: Amazon started blocking Muse within two weeks of its launch, so you can't tell it to buy things for you there. I am sure Amazon is working on their own agent internally. Apple is reportedly also working on a much needed improvement for an agentic Siri. And the domain dot.com, which changed hands in July, now redirects to xAI's Grok, further fueling the ongoing disagreement between Elon Musk and Sam Altman.

Why you care:

• Easy to try: Give them simple tasks: drafting replies from a read-only inbox, scheduling, supplier price lookups. Keep approvals on for anything that sends or pays.
• Keep it away from: your main bank account or card, payroll, customer lists and your password manager.
• Want this explained from zero? That is what Zero to Agents, above, is for!

President Trump recently said the US government is officially renaming Artificial Intelligence to Super Intelligence. Continuing on the push for SI, on September 29 Trump unveiled America.gov: an AI chatbot that answers government questions and points you to the right agency website. The General Services Administration operates it, and it pulls together "29,000 different federal government websites." It runs on Google's Gemini and xAI's Grok, per Al Jazeera, though reports conflict on the exact models. The President's pitch: the government now "stands only at your service."

What it does today: answers questions and links out. Social Security, Medicare, passport renewal, voter registration, business licensing.

What it does not do yet: process transactions or submit forms. Online passport applications are coming first: Spectrum News reports December, while Fox News reports Rubio saying 2027. Benefits applications are slated for early 2027. IRS filing is excluded.

The documented criticisms:

• Cost. No outlet has published a development cost, an operating budget or contract values, per one summary.
• Shifting answers. One explainer reports that within hours of launch the chatbot went from answering election-integrity questions to declining political ones, with no explanation.

Why you care:

• Today it is a shortcut, not an agent. Use it to find the right office, then read the agency's own page before you act on a fee or a deadline.
• Passports are the first real feature. America.gov will reportedly make passport renewal easier for Americans via a natural language process.

I took this photo the evening of September 1, looking south toward the Santa Teresa crossing, and went looking for the news story behind it. There isn't one.

What I did find is that it's not new. Customs and Border Protection flies a surveillance balloon, called an aerostat, at the Santa Teresa Border Patrol Station, "just west of the Santa Teresa Port of Entry", right where my camera was pointing. The aerostat went up in September 2024 with daylight and infrared cameras, and in June 2025 KVIA reported no plans to remove it. If I have the right balloon, it has been up for about two years surveilling the border.

What is not public, in anything I could open: what it costs, how far its cameras see, who can pull the footage and how long it is kept.

Across the river, Torre Centinela opened September 9 in Juárez (Chihuahua state government) with US diplomats and Las Cruces officials in attendance. The state's numbers: a tower more than 115 meters tall, more than 7,700 cameras, 1,791 license-plate readers and anti-drone systems. La Verdad Juárez puts the platform at about 4.2 billion pesos over five years and the opening more than two years late.

Why you care:

• Last issue El Paso unplugged about 150 plate readers. Chihuahua just opened the command tower for a statewide network that it says includes 1,791. If your staff, trucks or customers cross, assume the trip is logged on at least one side.
• If your own shop runs cameras, Wi-Fi tracking or a plate reader, write down how long you keep the data and who can see it before a customer asks.

The state had set a public hearing for October 19 on Project Jupiter's air permit, with a decision deadline of November 23. The hearing date has since come off the calendar; more on that below.

On September 24 Bloomberg reported that Oracle, the tenant at the Santa Teresa data center, sent a force majeure notice (Albuquerque Journal) to STACK Infrastructure, the developer, which Blue Owl Capital owns (Searchlight New Mexico). Force majeure is contract language for something outside my control is keeping me from performing. The notice itself is not public.

The reported cause is power. Two pieces are unresolved, per ENR: the state air permit for the roughly 2.45-gigawatt fuel-cell plant, and the gas pipeline to feed it, which slipped from August 2026 to February 2027 and still lacks federal approval to build. What the notice could do, as reported: let Oracle delay some rent payments for up to three years (El Paso Matters) if power is not ready on time, and only if Oracle and Blue Owl agree that a force majeure event has happened.

What it does not mean, so far: Oracle says such notices "do not, by themselves, establish a project delay or change delivery expectations." STACK says its financial commitments are unchanged. No outlet reports a stop-work order, a late payment or a notice to any subcontractor. As of mid-September the site was 26% complete, with roughly 3,500 workers (Searchlight New Mexico).

The Supreme Court fight is over, for now. On September 17 the New Mexico Supreme Court unanimously denied both challenges (Albuquerque Journal) and lifted its pauses on the air-permit case and the project's water well. On October 2 the state named retired federal judge James Browning as hearing officer for the air permit. No new hearing date has been announced.

Why you care:

• If you are a subcontractor or a supplier, nothing published says payments have slowed.
• If your bid touches the power plant, the equipment itself waits; site prep does not. The state says the project can't "build, modify, or install the actual emissions equipment before the permit is issued."

The robot lost to the S&P 500, and I'm shutting it down for now. At its last session, Friday, October 9, the account stood at about $103.74, against $105.71 for $100 of VOO bought the day this started, July 27. That is +3.7% for the robot and +5.7% for the S&P 500.

The test I wrote down in advance says the same. At session 30, on October 6, I checked whether the stocks the robot ranked highest went on to do better than the ones it ranked lowest. They did not: the relationship came out slightly negative, and the rule I set in August retires the signal "regardless of P/L."

What happens now: this version is done, pending a rewrite. Next I will redesign the trading strategy myself instead of letting an agent pick it. The agent will just run my strategy 24/7. The goal of beating the S&P 500 continues soon.

What I got wrong last issue. The table said September 4 through 8 was a "bot error" with no orders. Not right. September 4 was a missed run (the agent hit its AI model's usage limit), then came a weekend and Labor Day, and September 8 was a full session with five orders. It sold MU, NOW, JPM and TSLA, the four stocks I listed as "open," before that issue reached you. I also wrote "no orders" on a September 3 line that was itself an order. And the failures were in the scheduling and data plumbing, not the trading rules. My mistakes.

Closed trades since then, plus what was still open at the last session:

Add up the closed trades and you get about fifty cents lost. COIN made $3.55 over a weekend. Then on September 22 the robot put the whole account into SPXL, the 3x S&P fund, and sold it on the 28th when its market gauge flipped to caution: down $3.74. It led VOO as late as September 22 and trailed at every daily check-in after that. On October 5 and 6 it went back into SPXL, which is what it held at its last session. It still broke, too: on September 17 it reached the order step 28 seconds past its own cutoff and skipped a buy.

Every order since September 8:

Orders from before September 8 are on issue #03's scoreboard.

Why you care: an agent left alone to design its own trading strategy is not smart enough to beat the S&P yet. I found the trading agent to simply be too conservative in general, which could be due to how Anthropic post-trains its models for safety. In your business, be careful with handing financial tasks to agents.

None of this is financial advice.

That's issue #04. Zero to Agents is on the way, three companies want to give you an agent, Washington has a chatbot, Oracle has a legal notice out, and the robot lost to the S&P.

If you learned something here, forward it to one person. That's still how this list grows.

Coming up

• Project Jupiter's air-permit hearing.
• The first Zero to Agents episodes.
• Meta's ad blitz and the Borderplex AI-server corridor.

As always, thank you for the support,
Liberato and the Aguilabs team 🦅