Over the last few weeks, I have been writing about personal AI agents. Not the fancy version people show in demos, but the useful version normal professionals can actually use. An agent that can read something, think through it, take the next step, and save us from doing the same boring work again and again.

This week, OpenAI pushed that idea forward with three releases: ChatGPT-5.5, the new Codex experience, and ChatGPT Images 2.0. At first, this looks like another AI product update. New model. New features. New benchmark charts.

But I think this one is different. Let me break each one down and be straight about what actually matters for someone running agents day to day.

One Topic: Release of GPT-5.5, Codex Desktop and GPT Images 2.0

GPT-5.5: The Agent-First Model

Most people will ask the obvious question. Is ChatGPT 5.5 better than the last version?

Yes, it looks better. It is stronger at coding, research, data analysis, documents, and longer tasks. But that is not the most interesting part.

The real change is that ChatGPT-5.5 feels more useful for agent style work. It can understand a messy goal, plan the steps, use tools, check the result, and keep improving the output.

That matters because our daily work is messy or it also means you didn’t to write sophisticated prompts rather give brain dump, like:

“Can you check this and turn it into something useful?” That is where better agents win. Not by giving a clever answer, but by moving the work forward.

It is available as ChatGPT-5.5 & ChatGPT-5.5 Pro, also it’s expensive if run as an API. However, OpenAI claims that it should reason better and complete task in one or less iteration compare to previous model.

GPT Images 2.0: Text That Finally Works

The old GPT Images 1.5 model has one major problem. It generates good images about 40% of the time, but words inside images come out garbled or wrong. GPT Images 2.0 fixes this. Accuracy is now at ~99%, and to be honest, in my tests it gives much better output than the Nano Banana Pro and Nano Banana 2 models.

Other changes worth knowing:

• Resolution also increase upto 4K.
• Subject-lock editing lets you feed in a reference photo and keep that person, product, or logo consistent across different backgrounds and scenes.
• Multilingual text Japanese, Korean, Hindi, Arabic now renders natively inside images.
• And when you use it in Thinking mode, it reasons through complex visual requests like diagrams or infographics before generating anything.

For agent setups that produce social graphics, branded content, or weekly visuals this is the upgrade that was missing.

Codex Desktop: Why It Matters More Than ChatGPT

Here’s the part most people skip: ChatGPT is a conversation. You open it, ask, get a response, and move on. No persistent memory, limited file access, no persistent context, and the big concern is that every session starts blank. It fixes this to a certain extent by using Projects, though.

Codex Desktop is different. It remembers your projects. It can run tasks in the background while you’re doing other work. It can see and interact with your computer not just reply to text.

Think of it this way: ChatGPT is a smart advisor. Codex is a capable colleague who actually does the work and picks up where they left off. Codex is better when the work lives inside folders, files, apps, websites, and repeated steps.

For anyone building personal agents, that persistent, project-aware environment is what ChatGPT was never designed to be. That’s the real switch worth making. And with the way Codex development has progressed over the last few months, it looks like it will become a super app that brings together everything OpenAI will offer: chat, code, image, video, local computer use, and more.

GPT-5.5 vs Claude Opus 4.7

Claude Opus 4.7 launched April 16. GPT-5.5 dropped April 23. Back-to-back frontier releases the tightest competitive window in AI so far this year.

I think it’s very early to compare but here are some observation based on some initial test and published benchmarks, and the better question is: “Which model is best for this job?”

GPT-5.5 wins on: If you need an AI to navigate a website, run code, or control your computer, GPT-5.5 wins. It scored 82.7% on terminal-based tasks, while Claude hit 69.4%.

Claude Opus 4.7 wins on: For writing legal briefs, scientific papers, or high-level analysis that a human needs to check, Claude is still the gold standard. Claude still feels excellent when the task needs careful writing, review, judgment, and a more human sounding response.

So my view is simple. Use ChatGPT-5.5 when you want the AI to do more. Use Claude when you want the AI to think with you more carefully. If you are running personal agents, ChatGPT-5.5 may become the better engine for action-based workflows. But Claude should still stay in your toolbox. The smart move is multi-model routing: agentic tasks go to GPT-5.5, complex coding goes to Claude Opus 4.7, and simple, repeatable tasks go to cheaper models like Haiku or GPT-5.4 Instant. That is what the most efficient agent setups look like right now.

What This Means For Personal Agents

In the last few JustDraft editions, we talked about personal agents like OpenClaw, Hermes, and similar tools. The point was not to create a shiny AI toy. The point was to build a small system that helps with real life and real work.

This week’s releases make that idea stronger.

A personal agent can now do more than summarise your inbox or write a draft. It can research, create, compare, build, test, and improve.

But there is a catch. The more power we give agents, the more rules they need. I would not allow any AI agent to send emails, publish content, delete files, spend money, or message people without approval (Yet).

Powerful tools need boundaries.

Three Things to Try This Weekend

1. Codex Desktop – Start a real project. Open Codex Desktop, create a project folder, give it context about your work or business. See the difference from starting a blank ChatGPT window. The project memory alone changes what’s possible.

2. GPT Image 2 – Make a visual with text in it. Ask for a LinkedIn banner, a product mockup, or an infographic with actual words in it. Compare the result to what other model produce. The gap is significant.

3. GPT-5.5 Thinking – Watch the trace. Use it for a multi-step task like research brief, content plan, business analysis. Watch how it maps steps before acting. Redirect it mid-thought if the plan looks off. That interaction is the agentic experience OpenAI has been building toward.

Interested in travel or photography, read last week’s LensLetter newsletter about one lens walk around kit.

Read last week’s JustDraft about comparison between 10 major updates by OpenClaw in make your personal AI agent smarter.

Two Quotes to Inspire

When tools become powerful, judgment becomes the real advantage.

Agents don’t replace your thinking. They remove the distance between your idea and the output. What you do with that distance is still entirely on you.

One Passage Summary From My Bookshelf

Multipliers are leaders who use their intelligence to amplify the smarts and capabilities of the people around them. When these leaders walk into a room, lightbulbs go off over people’s heads; ideas flow and problems get solved. These leaders inspire employees to stretch themselves to deliver results that surpass expectations.

Diminishers are different. They are absorbed in their own intelligence, stifle others, and deplete the organization of crucial intellectual capital. In a world of rapidly advancing technology, the ability to be a Multiplier to harness the collective intelligence of both humans and AI is the ultimate competitive advantage.

From Multipliers: How the Best Leaders Make Everyone Smarter by Liz Wiseman