Building Your First Agent: From Chatbots to Autonomy
Move beyond AI chat into agentic systems that can read files, automate your browser, and run real workflows.
OpenAI Codex
Today's Agenda
What We'll Cover
01
Chatbot vs. Agent
Understand the core difference and why it matters
02
The 4 Pillars
What actually makes an agent an agent
03
Workspace Setup
Projects, permissions, memory files
04
Live Build
Morning Stand-Up agent — from scratch, live
The Chef Without a Kitchen
A chatbot is a world-class Michelin chef sitting in an empty room — brilliant mind, but no stove, no ingredients, no order ticket. They can only talk about cooking. An agent is that same chef, now standing in a fully equipped kitchen, ready to actually cook.
The Core Distinction
Same Brain. Completely Different Power.
🤖 Chatbot
Brain in a box. Brilliant reasoning locked inside a chat window. Can answer, advise, and explain — but cannot touch your world.
  • Responds to prompts
  • No access to files or apps
  • Output disappears after the session
🦾 Agent
Brain + Hands + Notebook + Destination. Same AI model, now equipped with tools, memory, and a goal to complete.
  • Reads and writes real files
  • Connects to your apps
  • Works toward a defined finish line
Framework
What Makes an Agent?
LLM — The Brain
The reasoning engine. Present in both chatbots and agents. This is the intelligence.
Tools — The Hands
APIs, browsers, and app integrations. This is what lets the agent touch your files and software.
Memory — The Notebook
Persistent files the agent reads each session so it never shows up empty-handed or forgetful.
Goals — The Destination
A specific outcome with a clear definition of "done" — not an endless open conversation.
The OODA Loop
This is the engine that keeps an agent working until it reaches its outcome. It loops continuously — observing, orienting, deciding, and acting — until the goal is achieved. The platform (Codex, Claude Code, etc.) is the harness that houses this engine.
Workspace Tour
No More "Floating Outputs"
Projects, Not Chats
In normal ChatGPT, deliverables end up scattered across random chats. Codex shifts your thinking to Projects.
Every project maps 1:1 to a real folder on your hard drive. Every file, document, and asset has a dedicated home — permanently.

Click "New Project" → use the file picker to bind Codex to a specific folder on your machine.
Keeping the Agent on a Leash
Default Permissions
Codex pauses before taking any major action and asks for your approval. Always start here. Monitor what it does before trusting it to run solo.
Full Access Mode
The agent runs completely autonomously without stopping to confirm. Only enable this once you've validated the workflow end-to-end.

Agents have "hands." They can take real actions. Always establish guardrails before granting full autonomy.
Memory System
The agents.md File
Start with Context - Before asking Codex to do any work, give it an onboarding document so you never repeat yourself. This is your agent's permanent rulebook.
Who you are
Your role, context, and background
Project goal
What this agent is built to accomplish
Your constraints
Hard rules the agent must always follow

Watch the context usage bar — feeding the agent too many files can overwhelm its memory window.
The Prompt Contract
Vague prompts cause agents to loop endlessly or invent their own finish line. Use a structured Prompt Contract every time.
1
Goal
The exact outcome — what does "done" actually look like?
2
Constraints
Strict guardrails. Tell the agent explicitly what it must not do.
3
Format
The exact shape of the deliverable — folder, doc type, layout. (you can give it reference documents)
4
Failure
What to do when stuck — "stop and ask," not loop and burn tokens.
Best Practice
Always Start in Plan Mode
Before unleashing the agent, toggle Codex into Plan Mode. In this mode, Codex will not execute any code or run automations.
  • Brainstorms a step-by-step approach
  • Surfaces clarifying questions early
  • Lets you approve architecture before any work begins

Review the plan → refine it → then flip the switch to execute.
Giving Your Agent Hands
To run our "Morning Stand-Up," the agent needs to connect to your actual apps. Here's what we'll wire up:
Gmail
Read the last 24 hours of emails — filter noise, surface what matters.
Google Calendar
Pull today's meetings and cross-reference with your priorities.
Slack
Deliver the final briefing as a clean, structured Slack message.
Step 1 — Project Setup
We’ll create a desktop folder named daily-standup and bind Codex to it as a new Project. Permissions stay on Default. Then we’ll initialize memory with agents.md.
Step 2: Create Agent.md
Speak don't type! (wisprflow.ai)
I want you to familiarize yourself with me and my work. I own a photography and media company and also do AI consulting look at the folder /jhr-website/ and learn about me. Get an idea for the business, and how I work, also look at the folder /jhr-vault/ for additional information.
Expect Codex will read the details and store this locally until you tell it to do something with it.
Now I want you to create an agents.md file forthis project. Thisfile will have information about me, as well as the goals and details of the project. Our ultimate goal is to create a daily brief that looks at my google calendar, my emails in the last 24 hours, and sends me a quick prioritiezed slack notification message to prioritize my day. We will connect these apps, and will build a skill to repeat this task daily.

Step 3 — Plan Mode
We’ll submit our Prompt Contract while Plan Mode is ON to produce a proposed step-by-step workflow.
Step 4 - Agent Contract
Goal: Access my Gmail to review the last 24 hours of emails, read my Google Calendar for today's events, and cross-reference these with my business goals documents. Create a daily plan, send it to me as a Slack message, and please propose a step-by-step plan for how you will execute this.

Constraints: Weed out all noise. Ignore promotional emails and spam. Do not send any emails on my behalf.

Format: A Slack message broken into three bulleted sections: 'Calendar Overview', 'Urgent Emails', and 'Action Plan for the Day'.

Failure: If an email is ambiguous or a meeting's purpose is unclear, do not guess. Stop and flag it under a 'Needs Your Input' section.
Codex should return a full plan with tool calls, filters, error handling, and Slack formatting.
Step 5 — Execute
With a solid plan, we toggle OFF Plan Mode and let Codex run. We’ll watch logs for skill execution and Slack delivery.

Test your app locally first. Make sure the output is what you need
Handling a Snag — Document and Fix
Assume Codex misses an email or mislabels it. We immediately capture the mistake and create a rule that the agents.md can follow so it never repeats, then re-run.
Root cause
Gmail “Newsletters” tag not excluded in filtering.
Permanent rule
Always filter out the “Newsletters” tag before ranking urgency.
Verification
Re-run with debug logs confirming exclusion.
Step 6: Iterate
Create a web dashboard display of my daily tracker that shows all the key details and activities for the day. Use the design.md skill and the UX/UI skill to make it sleek and modern and easy to view everything in one snapshot. This dashboard will reset each day with a new look so there's no reason to keep persistent memory.
Set the Timer
Turn this daily update into a skill that can be used multiple times and will run every day at 7 a.m. except for Sundays.
What the Briefing Looks Like
Your Slack message arrives each weekday morning:
Calendar Overview
Today’s meetings with times, locations, and prep notes.
Urgent Emails
High-priority items from the last 24 hours, de-duplicated.
Action Plan for the Day
Top tasks with owners and deadlines; “Needs Your Input” if blocked.
From Localhost to Live App
Codex builds → GitHub stores → Vercel publishes
1
Step 1 — Build & Test Locally
Verify everything works before sharing
2
Step 2 — Understand the Limitation
A localhost link cannot be shared
3
Step 3 — Create a GitHub Account
Use free GitHub account to store files.
Step 4 — Connect Vercel to GitHub
Sign into Vercel (Free) link it to GitHub
Step 5 — Share the Public URL
The app is now accessible
Step 6 — Update Locally then Push
Use Codex to make changes
The best part is if you get stuck - ASK CODEX!
Not Sure how to start your AI Agent Engine?

The Trust Engine Assessment gives you a personalized roadmap for integrating AI into your business.

Specifically for business owners who are ready to scale without losing the relationships that matter.
Jayson Rivas
Impact Consulting 931

Jayson@impactconsulting931.com
LinkedIn: /jayson-rivas/
Instagram: @jhrphotos