The Truth About AI Agents: They’re Multipliers, Not Replacements

Dr. Ryan Ries here. I have a couple of interesting AI developments to talk about this week, but first, I need to clear something up from the last Matrix.

Let Me Be Clear About Kiro

Last week I got pretty critical of AWS Kiro. My coworkers even reached out to me saying so.

I realized I need to provide more context.

I actually really like Kiro.

Kiro does something most AI coding tools don't. It forces you to think through your requirements before writing a single line of code. You start with specs, design documents, and task lists instead of jumping straight to code generation. 

This spec-driven approach solves a massive problem in AI-assisted development. When you come back to a project two months later, or hand it to a teammate, all the context is there. The prompts that led to the code. The decisions that shaped the architecture. The reasoning behind specific implementations.

Most AI coding tools give you vibe coding. You type a prompt, get code, maybe it works. Two months later, nobody remembers why anything exists.

Kiro gives you structure. It breaks down your natural language requests into formal specifications. It creates design documents. It generates task lists. Then it writes code based on those specs. The new property-based testing feature actually verifies whether the generated code matches your specification.

The checkpointing system lets you rewind changes when an agent goes sideways. The steering files give Kiro persistent knowledge about your coding conventions and preferences.

When I said things fell apart last week, I wasn’t specifically pointing a finger at Kiro. That's the reality of all current AI coding platforms. They're productivity multipliers for experienced developers, not replacements for engineering knowledge.

So yes, my SOW generation project didn't work perfectly. But the spec-driven approach Kiro is structured on meant I at least understood where things broke and why, and how I could fix it.

AWS is taking a very thoughtful approach to AI tools, like Kiro, which I appreciate.

OpenAI Thinks You Want AI in Your Group Chats

Now… here’s an AI feature that I am skeptical about.

OpenAI is testing a new feature that lets you add ChatGPT to group conversations with your friends, family, and coworkers. The pilot is running in Japan, New Zealand, South Korea, and Taiwan right now.

Up to 20 people can chat together with ChatGPT in the same thread. The AI is powered by GPT-5.1 Auto, which knows when to stay quiet and when to respond. You can also summon it by typing "ChatGPT" in the conversation.

OpenAI built in privacy safeguards. Your personal ChatGPT memory doesn't transfer to group chats. If someone under 18 joins, the AI automatically reduces sensitive content for everyone.

The obvious use cases: 

• trip planning
• event coordination
• study groups
• work projects 

…where everyone needs to drop links and files while ChatGPT searches, summarizes, and generates content on top of that shared information.

Here's what I want to know… do people actually want this?

Most users prefer one-on-one interactions with AI but perhaps that is just because right now that’s all most users have experienced. 

Here’s where I see a big problem, though. Adding a bot to your group chat means everything you say becomes training data for OpenAI's future models. They promise to strip user-identifying information first, but it still feels too invasive.

OpenAI is betting that convenience outweighs privacy concerns. They might be right. If people start adding non-ChatGPT users to these group chats just to collaborate, that's a lot of new users for OpenAI.

But I'm skeptical. This feels like a feature OpenAI wants us to want, not something people are demanding. We'll see if it expands beyond the pilot regions.

The Human-Agent Productivity Study You Need to See

Upwork just released results from their Human-Agent Productivity Index (HAPI), one of the first real evaluations of AI agent performance on actual client work.

They tested AI agents from Claude Sonnet 4, Gemini 2.5 Pro, and OpenAI GPT-5 across 300 real Upwork projects. They intentionally chose simple, well-defined tasks where agents had a reasonable chance of success.

The Results Are In

AI agents perform well on tasks with objectively correct answers like math or basic coding. With qualitative work, such as landing page design, marketing copy, and strategic analysis, agents struggled without human guidance.

I thought this was great because it doubles down on what we have all been saying. 

When human experts collaborated with agents, job completion rates jumped 70% compared to agents working alone. This pattern held across different types of work.

Upwork CTO Andrew Rabinovich points out that "where a project might take a freelancer days to complete independently, the agent-plus-human approach can deliver results in hours through iterative cycles of automated work and expert refinement."

This matches exactly what I experienced with Kiro last week. The agent can move fast, but you need human expertise to direct that speed toward useful outcomes.

The study tested agents on simple projects that represent less than 6% of Upwork's total work volume. These are tasks AI should theoretically handle well. Even there, human expertise made all the difference between success and failure.

My Thoughts

Three stories this week, one consistent theme: AI tools work best when they acknowledge their limitations and design for human collaboration.

Kiro builds structure and documentation into the coding process because unstructured AI code becomes unmaintainable. 

OpenAI is testing group chats that might solve collaboration problems or might just create new privacy concerns. 

Upwork's study confirms that agents plus humans dramatically outperform agents alone, even on simple tasks.

What does this tell us? 

We're not at the "AI replaces humans" stage. We're at the "AI amplifies human expertise" stage. The companies and individuals who succeed will be those who figure out how to effectively direct AI capability rather than those who try to fully automate away human judgment.

What are you seeing in your work? Are you finding effective ways to collaborate with AI agents?

Until next time,
Ryan

Now, time for this week’s AI-generated image and the prompt I used to create it:

Create an image of a muppet in a Mission, a CDW company t-shirt. The Muppet is wandering through a scene in the movie Wicked. You can see Glinda, played by Ariana Grande, and Elphaba, played by Cynthia Erivo, in the background.