A free video course from Aguilabs

Zero to Agents

Go from your first AI conversation to practical agents that can help with real work and everyday tasks.

Made for everyday people. No tech knowledge needed: every idea is explained from zero.

Subscribe to Aguilabs SignalSee the episodes

Each episode is announced in Aguilabs Signal, our free newsletter.

MTWTFSSBreakfastLunchDinnerCaloriesProtein
The tracker. Every episode adds one feature to it.

In production. No episodes are published yet.

Price
Free
Languages
English and Spanish
Length
18 episodes planned, in 5 parts
Format
Animated video, narrated by the founder of Aguilabs

One app, built on screen

Every episode works on the same small app: a meal-prep tracker that plans the week's meals, logs what you ate, counts calories and protein, and writes the grocery list. What changes is who does the building. Each episode hands a little more of the work to the machine.

  • The machine does it
  • Still done by hand

  1. LLM

    1 of 7 jobs handed to the machine

    Writes the text: a recipe, a plan, a piece of code.

    Episode 1

  2. Chatbot

    2 of 7 jobs handed to the machine

    Keeps the conversation going, so you can ask for changes.

    Episode 2

  3. Tools

    3 of 7 jobs handed to the machine

    Looks up real data and uses other programs, instead of guessing.

    Episode 9

  4. Agent

    4 of 7 jobs handed to the machine

    Edits the files and runs the tests itself. The copy and paste ends.

    Episode 11

  5. Subagents

    5 of 7 jobs handed to the machine

    Splits a big job into pieces and hands them out.

    Episode 14

  6. Swarms

    6 of 7 jobs handed to the machine

    Works as a team of agents, each with its own task.

    Episode 16

  7. Autonomous

    7 of 7 jobs handed to the machine

    Takes a request and ships the feature, with a person checking the work.

    Episode 17

Where it starts

You copy code out of a chat window and paste it into files, one block at a time. You are the model's hands.

Where it ends

Someone asks for a new feature from their phone, and the software builds it, tests it and ships it, with a person approving the steps that matter.

The episodes

This is the plan as it stands today. Episodes may be added, merged or reordered while the course is being made, and there are no release dates yet.

Published
0
In production
2
Planned
16

Next upEpisode 1: What is an LLM?

Episode 0

Introduction

  1. Episode 0: Start here

    A short introduction: who teaches the course, why it was made, what the episodes cover, and the finished tracker.

    In production

Part 1

The model

What an AI model is, what it can see, and how to ask it for something useful.

  1. Episode 1: What is an LLM?

    How an AI model writes: one word at a time, by predicting what comes next. Why it can sound sure and still be wrong. In the tracker, it writes our first function.

    1. Guess what comes next
    2. Tokens, not words
    3. A list of probabilities, then a pick
    4. The generation loop
    5. Inside the box, and where its numbers come from
    6. What the model can see
    7. Plausible is not the same as true
    8. It is only text

    In production

  2. Episode 2: The chatbot

    The app wrapped around the model: messages, roles and a hidden first instruction. We build the first version of the tracker by copy and paste, with you as the model's hands.

    1. The model keeps nothing
    2. A chat is one growing document
    3. The hidden first message
    4. The window fills up
    5. What makes a request good
    6. Version zero, by copy and paste
    7. The error loop
    8. What a chatbot cannot do, and the road so far

    In production

  3. Episode 3: Prompts and context

    What you put in front of the model matters more than clever wording: instructions, examples, files. We write the spec for the tracker.

    1. Same model, two requests
    2. A prompt is everything in the window
    3. Each part removes a guess
    4. Show, don’t tell
    5. Paste the real files, but not everything
    6. From a wish to a spec
    7. Plan first, and keep your drafts
    8. The tracker, version one

    Planned

  4. Episode 4: Choosing a model

    Big or small, fast or careful, online or on your own computer. How to read the comparisons with care and pick the right model for each job.

    1. Not one model: a shelf
    2. Three dials, tied together, and three sizes
    3. Tokens are the bill
    4. Thinking before answering
    5. Windows come in sizes
    6. Hosted or local
    7. The test that matters
    8. Two jobs, two models

    Planned

Part 2

More than text

Models that see, speak, draw and search.

  1. Episode 5: The model can see

    Images and documents as input. A photo of a plate becomes an estimate of what is on it.

    1. The same model, a new kind of input
    2. How a photo becomes something the model can read
    3. What it does well
    4. Where it stops
    5. Images take up window space
    6. Documents are text, pictures or both
    7. A photo is evidence, not a measurement
    8. The tracker: photo in, estimate out, you confirm

    Planned

  2. Episode 6: Generating images and video

    How image generation works, in one picture, and what it is good and bad at. The tracker gets its icons and recipe photos.

    1. A different kind of model
    2. Diffusion in one honest picture
    3. The seed and the prompt
    4. Why hands, lettering and consistency are hard
    5. Editing is not generating
    6. Video: images that must agree over time
    7. Rights, likeness and labelling
    8. The tracker gets its pictures

    Planned

  3. Episode 7: Voice

    Speech to text, text to speech and live conversation. Log a meal by saying it out loud.

    1. Three pieces, not one
    2. Speech-to-text: sound becomes words
    3. Where transcripts go wrong: accents, noise and names
    4. Text-to-speech: words become sound
    5. Latency and turn-taking
    6. Real-time speech-to-speech
    7. Voices that clone: consent and disclosure
    8. The tracker listens and reads back

    Planned

  4. Episode 8: Embeddings, search and classifiers

    How a model turns meaning into coordinates, so the tracker can search a recipe library by meaning and tell a meal from a question.

    1. Meaning as coordinates
    2. Nearest-neighbour search
    3. Retrieval: fetch, then put it in the window
    4. Chunking, and why retrieval can miss
    5. When a plain keyword search is the better tool
    6. Classifiers and routers: small, cheap decisions
    7. Guardrails are classifiers too
    8. The tracker: a library you can search, and a sorter

    Planned

Part 3

Giving it hands

The model stops describing the work and starts doing it.

  1. Episode 9: Tools and structured output

    The model asks for a tool and uses the answer. The tracker looks up real nutrition data instead of guessing.

    1. The model can only write
    2. A tool is a program plus a description
    3. The tool-call loop
    4. Structured output: an agreed shape
    5. The description is the instruction manual
    6. Results are text, and tools can fail
    7. The tracker gets real numbers

    Planned

  2. Episode 10: The command line and MCP

    The terminal as the tool that reaches everything, and MCP as a standard plug for the rest. We connect a database and a grocery service.

    1. One tool at a time does not scale
    2. The command line: the computer as a conversation in text
    3. Why a model is at home there
    4. MCP: a standard plug
    5. Hosts, clients and servers
    6. Local and remote
    7. Every connection widens the reach
    8. The tracker gets connected

    Planned

  3. Episode 11: The agent

    A model, its tools and a loop. It edits the files and runs the tests itself, within the permissions you give it. The copy and paste ends.

    1. From one call to a loop that keeps going
    2. A coding agent at work
    3. The copy and paste ends
    4. A plan first, and a way to check
    5. When it gets stuck
    6. Permissions and approval prompts
    7. What you review now
    8. The tracker builds its own feature

    Planned

  4. Episode 12: Harnesses

    Everything wrapped around the model: instructions, memory, permissions. Why the same model behaves differently in different products.

    1. The same model, two products
    2. The system prompt and the tool set
    3. Managing the window
    4. The project’s instruction file
    5. Asked or enforced: permissions and hooks
    6. What you can set, and what you cannot
    7. The tracker gets its rules

    Planned

  5. Episode 13: Skills and reusable prompts

    Skills, commands and playbooks: teaching an agent how you work, once, so you do not have to repeat yourself.

    1. You keep explaining the same thing
    2. Reusable prompts and commands
    3. Why not put it all in the instruction file?
    4. Skills: loaded only when needed
    5. Writing one
    6. Skill, instruction file or prompt?
    7. Sharing skills
    8. The tracker’s way of adding a feature

    Planned

Part 4

Many agents

One agent becomes a team, and the team checks its own work.

  1. Episode 14: Subagents

    Delegation. One agent hands pieces of the job to others that explore, plan, build and review, each starting with a clean slate.

    1. The window fills up again
    2. A second agent with a clean window
    3. Four roles: explore, plan, build, review
    4. What is lost in a hand-off
    5. Writing a brief another agent can work from
    6. When one agent is enough
    7. The tracker gets a weekly report

    Planned

  2. Episode 15: Orchestration

    Workflows, parallel work and checking loops, with tests as the contract that says when the work is done.

    1. From a conversation to a procedure
    2. Parallel branches
    3. Isolation: a copy of the project each
    4. The verification loop
    5. Tests and evaluations: the contract for done
    6. An independent reviewer
    7. Retries and stopping rules
    8. Two features at once in the tracker

    Planned

  3. Episode 16: Swarms

    Teams of agents: when more agents help, when they do not, and what it costs.

    1. From two agents to many
    2. Wide work: where more agents help
    3. Coupled work: where they do not
    4. Coordination cost
    5. Agreement and voting
    6. Errors that multiply
    7. What it costs
    8. The tracker’s new look, and where things stand

    Planned

Part 5

Autonomy

From a request to a shipped feature, and how to keep that safe.

  1. Episode 17: Autonomous development

    From a request to a shipped feature, with human checkpoints. Someone asks for a tracker feature from their phone, and it ships.

    1. The whole line
    2. A request is not a spec
    3. Checkpoint one: approving the spec
    4. Build, review, test
    5. Checkpoint three: the release
    6. What autonomous should mean
    7. Who is accountable
    8. From a phone to the tracker

    Planned

  2. Episode 18: Safety, cost and what is next

    What can go wrong and how to contain it: tricked models, sandboxes, budgets and tests. And where all of this is heading.

    1. Instructions and data are the same text
    2. The recipe that gives orders
    3. Least privilege
    4. Sandboxes and secrets
    5. Approvals
    6. Budgets and runaway loops
    7. Evaluate before you trust
    8. The tracker holds, and what is next

    Planned

Finish the course and your first month is free

Pass the quiz of all 18 episodes and the first month of the Starter plan is free when you become an Aguilabs client.

The rules

  • The first month of the Starter plan is free.
  • It is for new clients only.
  • One per business.
  • Claim it within 90 days of finishing the course.
  • It has no cash value.
  • It cannot be combined with another offer.

You claim it on this page once you finish, with the account you took the course with. Nothing is charged and no card is asked for. We write to you, and the free month is applied by hand when your project starts.

See what the Starter plan includes

Follow the course

Each episode is announced in Aguilabs Signal, our free newsletter on practical AI and border tech, in English and Spanish. Subscribe, and you will know when the first one is out.

Subscribe to Aguilabs SignalRead past issues of Aguilabs Signal

Or follow Aguilabs on