AI Builders Summit

Codex for Everyone

Aishwarya Naresh Reganti + Kiriti Badam · August 17, 2026

LevelUp Labs

Meet Your Hosts

Agent builders and practitioners helping teams become AI-native while keeping humans at the center.

Aishwarya Reganti

Co-founder · CEO

Applied AI researcher, former AWS leader, and instructor at MIT and Oxford.

Kiriti Badam

Key engineer behind OpenAI Codex, with 10+ years building large-scale systems and ML infrastructure.

AI Builders Summit

A free online conference for AI builders, by AI builders.

August 8 – September 15, 2026 · Live sessions with the people shipping AI in production.

summit.levelup-labs.ai ↗

LevelUp Labs paper mascots gathered around a table — Source: LevelUp Labs · AI Builders Summit · 2026

The Summit Keeps Growing

Catch up on completed talks. Join the next sessions live.

summit.levelup-labs.ai ↗

**Zoë Barnswell**AI builder, Wharton MBA

**Chris Alexiuk**Product Research Engineer

**Ed Charbeneau**Developer Relations Lead

The basic mechanism

An Agent Is a Loop

A model calls tools, observes the result, and repeats until the task is complete.

Agent loop where a model calls tools and uses their observations before producing a result — Source: LangChain · Agents documentation · 2026

What “Codex” means

Codex Has Three Layers

01Model

The intelligence that reasons about the task.

02Agent system

Context, tools, permissions, and the execution loop.

03Surface

The app, CLI, IDE, web, mobile, and remote interfaces.

2025 · Cloud agent

Codex Started as a Cloud Agent

The cloud agent added isolated environments, parallel tasks, tools, and iterative verification.

Introducing Codex ↗

Codex cloud environment configuration with an isolated terminal — The model gained an execution environment and a way to check its own progress.Source: OpenAI · Introducing Codex · 2025

2025 · Developer tools

Codex Entered the Developer Workflow

Codex CLI running in a terminal — **CLI**Source: OpenAI · Codex CLI · 2025

Codex working inside an IDE — **IDE**Source: OpenAI · IDE extension · 2025

A GitHub comment triggering a Codex code review — **GitHub**Source: OpenAI · Code review · 2025

Slack

Codex Can Start From Slack

Mention Codex in Slack. The conversation becomes context for a new task.

“If you put Codex inside Slack, it can figure out what it needs as context.”Kiriti Badam · OpenAI, via Slack

Read the Slack customer story ↗

Codex being assigned work from a Slack conversation — Source: OpenAI + Slack · Codex in Slack · 2026

Remote

Remote Keeps You in the Loop

From your phone, start, guide, approve, and review work running on a connected computer.

The connected computer must remain awake and online.

Codex task list on a phone — Source: OpenAI · Codex remote access · 2026

2026 · Desktop app

The Codex App Brings Agents Together

Projects, parallel tasks, review, skills, and scheduled work came together in one place.

Codex app launch ↗

Codex desktop app with projects, a task thread, and the review pane — One place to delegate, follow progress, and review the result.Source: OpenAI · Introducing the Codex app · 2026

How Codex Knows Your Work

01Context

Projects supply files and instructions.

02Tools

MCP connects systems and data.

03Methods

Skills and plugins make expertise reusable.

Project context

Projects Supply the Working Context

Open a folder and Codex can inspect its files, including the instructions that define how work should happen.

agent-native-operator/ ↳ AGENTS.md ↳ lightning-session/ ↳ course-materials/ ↳ skills/

AGENTS.md tells Codex how to work in this project.

Tools

MCP Servers Give Codex Tools

An MCP server gives Codex named tools for reading data or taking actions.

Codexchooses a tool

→

MCP serverhandles access and returns structured results

→

Search mailRead calendarCreate draft

Tools define access; your task defines authority.

Expertise

Skills Load Expertise on Demand

1

Always visibleSkill name + description

↓

2

Task matchesCodex selects the skill

↓

3

Loaded on demandFull instructions and resources

Actual skill used for this deck

---
name: levelup-labs-html-presentations
description: Build and verify LevelUp Labs HTML decks.
---

# Presentation skill

- Inspect the source and assets
- Build
- Verify

Packages

Plugins Bundle Capabilities

A plugin is an installable package that can combine expertise and connections for a particular kind of work.

Plugin

SkillsHow to do the work

Connectors + MCPWhich systems Codex can reach

Optional resourcesScripts, assets, templates, and hooks

June 2026

Codex Comes With Many Plugins

It also integrates with a very large range of services.

6role-specific plugins

62popular apps

110skills

Data analyticsCreative productionSalesProduct designPublic equityInvestment banking

OpenAI launch ↗

Demo 1

Connect Gmail and Calendar

Tools provide access. The prompt sets boundaries.

GmailGoogle Calendar

Check today’s Calendar and 24 hours of Gmail. What needs my attention, and what should I prepare for?

Prioritize urgent requests and meeting context. Link every source.
Don’t send or change anything.

AI-native artifacts

HTML Is the Presentation Layer

PPTX is optimized for office editing. HTML exposes the structure, styling, and rendered result directly to both people and agents.

PPTXManual slide editing

→

HTMLReadable · editable · previewable · verifiable

Quick demo

Build One Slide From Scratch

Start from a blank task.

Create one 16:9 HTML slide titled “The Future of Work” using a black background, large white type, and one cobalt accent.

Save it as a standalone index.html with no external dependencies, then open it in the browser.

Rendered work

Verify in the Browser

Codex can open a local site, inspect the rendered page, interact with it, and verify what a person would see.

Codex in-app browser showing a rendered webpage beside the task — The rendered result is the evidence.Source: OpenAI · Browser documentation · 2026

Focused iteration

Annotate the Exact Change

1Select

Point to the exact region.

→

2Comment

Say what should change.

→

3Revise

Codex updates that part in context.

Keep what already works. Refine the part that does not.

Demo 2

Create, Review, Verify

SkillLevelUp design system

→

HTMLBuild the deck

→

BrowserVerify the result

→

Approved deckReady to share

Sites

Sites Can Run Full Apps

Sites host websites, web apps, and games, with durable data and file storage when needed.

Hosted experienceCreate, deploy, and share a production URL.

Transactional databaseKeep records, progress, scores, and structured data.

Object storageStore images, documents, audio, video, and uploads.

Official Sites documentation ↗

Demo 2 · Publish

Publish the Deck as a Site

Turn the approved deck and its resources into a hosted page anyone with access can open.

Deck→Save version→Review→Deploy→Public URL

Advanced Codex Features

AppshotsSubagentsVoiceChatGPT WorkGoal mode

They add richer context, parallel work, hands-free control, cloud execution, and longer-running goals.

Appshots

Appshots Show Codex What You See

Capture the frontmost Mac app as an image, along with any text the app makes available.

You choose when to capture. Codex does not continuously watch the screen.

A Numbers budget spreadsheet captured as an Appshot — Useful when the work lives in another desktop app.Source: OpenAI · Appshots documentation · 2026

Parallel delegation

One Task Can Spawn Specialist Tasks

Parent taskOwns the outcome

Research sourcesAudit the copyFind visual assets

↓

One collected resultThe parent integrates the parallel work.

Hands-free coordination

Direct Work by Voice

StartOpen a new task

CheckAsk for progress

SteerChange the direction

Voice coordinates tasks. The same permissions and approval rules still apply.

Cloud work

ChatGPT Work

Cloud agents carry the work; people return for judgment.

No more keeping the laptop just slightly open to actually get work done.

LOCAL RITUALKeep the lid cracked.

CLOUD DEFAULTStart work. Close the lid.

Satirical OpenClaws product advertisement showing a small metal claw holding a MacBook slightly open so background agents can stay alive — Source: Reddit · “OpenClaws, an accessory for OpenClaw agents” · 2026 · original post ↗

Goal mode

Engineer the Longer Loop

For multi-step work, define the outcome, constraints, and how completion will be verified.

OutcomeWhat must change?

ConstraintsWhat must remain true?

VerificationWhat proves completion?

→/goal

The basic agent loop stays the same. The completion contract becomes explicit.

Demonstration 3 · Automation

Automate Only After It Works

1Run once

Is the result useful?

2Review

Fix scope and boundaries.

3Schedule

Preserve the working contract.

4Monitor

Keep results reviewable.

Attention systems

Turn Repeated Signals Into Reviewable Briefs

Email

What needs my attention today?

Personal priorities

Product feedback

What themes keep recurring?

Product decisions

Course Slack

What questions remain unanswered?

Teaching gaps

Beyond these examples

More Work You Can Delegate

UnderstandResearchMeeting prepOperating reviews

PrioritizeDaily briefsInbox triageFeedback radar

CreatePresentationsSitesInternal tools

OperateScheduled checksReview queuesTeam workflows

We’ve Only Scratched the Surface

Features are the beginning. Operating them well is the craft.

From Features to Operators

Now we decide what Codex may do, how it knows it is finished, and what a human checks.

ContextAuthorityDefinition of doneVerificationFeedback

Operator reliability

Design a Trustworthy Operator

Set authority→Define “done”→Verify the result→Retain feedback↺

When confidence is low or authority is missing, the operator asks for help.

Operator craft

The Operator Layer

Operators design the system around the agent.

LLM wikiDurable memory

Cost-efficient orchestrationSpend compute intentionally

Self-improving loopsImprove from evidence

Feedback mechanismsCapture corrections

Skill creationEncode expertise and guardrails

And many more.

Saturday · August 22

Learn the Operator Craft

A hands-on session for turning impressive agent capabilities into dependable ways of working.

Saturday August 22, 2026 10 am–4 pm PT LevelUp Labs · live working session

QR code linking to the Saturday Operator workshop — Scan for the workshop

Your turn

What Would You Delegate?

Choose one task that repeats, has clear evidence, and ends in work you can review.

Questions

All courses and resources

Maven profile ↗