How Two AI Agents Work Together Without Code (Non-Developer Setup) | Practical AI Ep 34

Practical AI: Episode 34

I’m Not a Developer. Here’s How I Got Two AI Agents to Work Together.

Watch on YouTube

Published: March 27, 2026

TL;DR

  • OpenAI had its worst week ever. Sora is dead, the $1 billion Disney deal evaporated, and Apple is opening Siri to all AI models. Then they raised $10 billion anyway.
  • Power users build systems, not answers. Anthropic’s Economic Index shows power users hit 73% task success by iterating and building reusable workflows. Novices accept first results and stall at 67%.
  • AI amplifies whatever you already are. 81,000 people interviewed about AI. Entrepreneurs extract the most value (47%). Employees with side projects outperform everyone (58%). The gap is not technology. It is intent.
  • Two AI agents can work together without code. Olga demos her system: Claude Code (Athena) and OpenClaw (Knox) share a Dropbox folder, trade session logs, and hand off work across four businesses. No GitHub. No development experience required.
  • AI funding hit $13.4 billion this week. OpenAI’s $10 billion accounts for most of it, but even without that round, funding doubled from the prior week.

Table of Contents

About This Show

Practical AI is a weekly live show (Fridays 11am CT) hosted by Olga Pechnenko and Chris Pearson that cuts through AI hype to deliver news, trends, and hands-on tips for builders and founders. Unlike technical AI podcasts, Practical AI focuses on business applications and ROI. What actually works, what is hype, and what you can implement Monday morning.

What You’ll Gain

  • Understand what power users do differently based on Anthropic’s analysis of millions of AI interactions. The gap between novice and expert is not about the tool. It is about whether you accept first results or iterate toward a reusable system.
  • Learn how to set up two AI agents that talk to each other using nothing more than a shared Dropbox folder and a simple session log file. No development experience needed. Olga walks through her exact setup running four businesses.
  • Discover why OpenAI lost three major partnerships in one week and still raised $10 billion. The Sora shutdown, the Disney deal collapse, and Apple opening Siri to competitors tell a story about where the AI industry is actually heading.
  • See where AI reasoning actually stands today with ARC AGI-3 benchmark results. Every frontier AI model scores below 1% on novel reasoning tasks. Humans score 100%. The gap between demos and real intelligence is wider than most people think.
  • Gain perspective on how 81,000 people across the world actually use AI. The dominant fear is not job replacement. It is trusting a system that hallucinates. And the people extracting the most value are entrepreneurs and employees with side projects.

Biggest Takeaway to Implement: Stop trying to get a single result from AI. Start building a system you can run a billion times. Power users spend their time refining workflows and reusable skills. Novices accept the first output and move on. The system is the product.

PageMotor and Practical AI Updates

Free, informative, and FUN!

Frequently Asked Questions

What happened to OpenAI’s Sora video tool?

OpenAI shut down Sora entirely. Consumer app, developer API, ChatGPT integration. All gone. The $1 billion Disney licensing deal died with it. OpenAI is redirecting resources toward AGI ahead of their 2026 IPO. Read more below.

What is the difference between AI power users and novices?

According to Anthropic’s Economic Index, power users achieve 73% task success versus 67% for novices using the same tools. The difference: power users iterate and converse instead of accepting first results. They build reusable workflows, not one-off outputs. Read more below.

How do you get two AI agents to talk to each other?

Olga uses a shared Dropbox folder with a SESSION-LOG.md file in each project. One agent writes what it did. The other reads it and picks up where it left off. No code, no GitHub, no technical setup. Read more below.

What is ARC AGI-3 and why does it matter?

ARC AGI-3 is a benchmark testing whether AI can reason through completely unfamiliar problems with no instructions. Every frontier AI model scores below 1%. Humans solve 100% on the first try. It is a reality check on how far AI reasoning still has to go. Read more below.

Who benefits most from AI right now?

Anthropic’s study of 81,000 people found that 47% of entrepreneurs extract the most value from AI, compared to just 14% of institutional employees. Employees with side projects hit 58%. Freedom to explore and apply AI without corporate constraints is the key factor. Read more below.

Is Apple replacing ChatGPT in Siri?

Apple plans to open Siri in iOS 27 so users can choose any AI model installed through the App Store. ChatGPT loses its exclusive role. Claude, Gemini, Grok, and others become options. Expected to ship fall 2026. Read more below.

How much AI funding was raised this week?

$13.4 billion in AI funding across 89 companies. OpenAI accounted for $10 billion. Even without that, funding doubled from the prior week. There were 11 mega rounds above $100 million. Read more below.

Practical AI Episode 34: I’m Not a Developer. Here’s How I Got Two AI Agents to Work Together.

Key Definitions

What is A2A (Agent-to-Agent)?

A2A describes a world where AI agents interact directly with other AI agents to complete business tasks. Instead of humans navigating websites and filling forms (B2B), agents hand off data to other agents in chains. Alibaba’s president described this as the next evolution beyond B2B. Your business needs to be ready for agents to find and interact with it.

What is a RAT (Repetitive Annoying Time-consuming Task)?

A framework from Lauren Goldstein (Episode 33 guest). RATs are the five-minute tasks that feel too small to fix but compound into a month of lost time per year. The best use of AI is not the big two-hour projects. It is eliminating these paper cuts one by one.

What is ARC AGI-3?

A benchmark from the ARC Prize Foundation that tests whether AI can reason through completely novel, unfamiliar problems with no instructions. Unlike standard benchmarks that test trained knowledge, ARC AGI-3 measures whether AI can explore, adapt, and figure things out from scratch. Current frontier models all score below 1%. Humans score 100%.

What is a CLAUDE.md file?

A plain text file placed in a project folder that gives an AI agent all the context it needs about that project. Who the people are, what has been built, what the current state is, how the user likes to work. Without it, every AI conversation starts from zero. With it, the agent already knows everything relevant to your business.

Quotable Moments

Don’t think about running it once. Think about running it a billion times. Then suddenly you’ll have a light bulb moment.

— Chris Pearson on why power users think differently about AI

AI does not create drive. It amplifies baseline intent.

— Olga Pechnenko on the 81,000-person study

There’s no drama. Zero ego. When Athena picks up Knox’s work, she doesn’t redo it. She builds on it. These two agents treat each other’s work with more respect than most human coworkers.

— Olga Pechnenko on her two-agent system

Good people remove the need to do a lot of that stuff. I think we’re just now starting to realize how much good personnel enabled us to be lazy about management and about context.

— Chris Pearson on what AI management reveals about human management

2:47 OpenAI’s Terrible, Horrible, No Good Week

Three blows landed on OpenAI in a single week. First, they killed Sora entirely. No consumer app, no developer API, no ChatGPT integration. The $1 billion Disney licensing deal that would have opened 200+ Disney, Marvel, Pixar, and Star Wars characters to AI video generation died with it. OpenAI is redirecting all those resources toward AGI ahead of their planned 2026 IPO.

The Sora Timeline

Sora launched, triggered massive copyright backlash within 24 hours (someone made South Park characters say things South Park never wrote), burned through compute on what Chris called “goofball slop videos,” signed the Disney deal to fix the copyright problem, then killed the whole thing six months later.

The next day, Apple announced it will open Siri to all AI models in iOS 27. ChatGPT loses its exclusive role as Siri’s AI partner. Users will choose whatever model they have installed from the App Store. Claude, Gemini, Grok. All options. Chris called this “Apple’s first smart move in the AI space.” Expected to ship fall 2026.

Key Takeaway: A trillion dollars in market cap walked away from OpenAI in one week. If the biggest deals in AI can evaporate overnight, build on tools that deliver ROI today, not moonshots.

7:28 The Infrastructure Layer: Chips, Search, and Agentic Hardware

Three infrastructure stories dominated the week. Elon Musk announced Terafab, a $25 billion chip factory in Austin, Texas. Tesla, SpaceX, and xAI will design, produce, and iterate their own chips domestically. The goal: 50x current chip production capacity, powered by solar energy from space-based data centers.

ARM’s Agentic CPU

ARM launched a 136-core CPU built specifically for parallel AI agent execution. All 136 cores operate simultaneously. The analogy Olga used: we are at the point of paving highways for AI, just like we paved roads when cars replaced horses.

Google rolled Search Live to 200 countries. Voice and camera overlay lets you show Google what you see and ask questions about it. Chris demonstrated the concept: walk into a shoe store, show the wall of shoes, ask “which ones have a wide toe box?” and Google tells you based on the companies’ existing online data. This is GEO in action. If your product information is not discoverable by AI, you are invisible.

Google also relaunched AI Studio with Firebase for vibe coding. Chris was skeptical. Firebase ties you to Google’s ecosystem. He compared it to Lovable’s approach and questioned why the multi-service model (database here, auth there, hosting somewhere else) is the default when a one-stop solution like PageMotor keeps everything on one server.

17:40 Claude Ships Auto Mode and Remote Mac Control

Anthropic shipped two features this week. Auto mode for Claude Code lets you turn off the constant permission prompts. No more “can I read this file?” every three seconds. As Olga put it: “I already give you permission. Use this folder. Just read the file.” Claude Cowork also gained remote Mac control. It can take screenshots, move the mouse, and execute actions while you are away.

Pro tip: Olga noted that Claude Code and Cowork have different strengths. Cowork can create Excel sheets. Claude Code cannot. For client files and spreadsheet work, she uses Cowork. For deep project work in the terminal, she uses Claude Code.

24:10 ARC AGI-3: The Reality Check on AI Reasoning

The ARC AGI-3 benchmark dropped this week and reset the scoreboard. These are abstract puzzle games with no instructions. The AI has to explore, figure out the hidden rules, and solve them from scratch. Every frontier model. Gemini, GPT, Claude, Grok. scored below 1%. Humans solve 100% on the first try.

The ARC Progression

ARC AGI-1: AI went from 30% to 93% in two years. ARC AGI-2: AI went from 25% to 68% in one year. ARC AGI-3: every model on the planet scores under 1%. Humans still 100%. There is a $2M prize to push progress.

Chris connected this to his real experience building with AI. Two-dimensional tasks (convert a homepage design) work well. Three-dimensional tasks (build an entire multi-page website with menus, team pages, landing pages) break down fast. He described being hundreds of reps in and still playing whack-a-mole with errors. The flashy demos work on familiar, well-trained tasks. On truly new problems, AI still cannot independently reason, explore, and plan.

30:48 What AI Power Users Do Differently

Anthropic released their Economic Index report analyzing millions of AI interactions. The data reveals four dimensions that separate power users from novices.

The Success Gap

New users (under 6 months) achieve 67% task success. Power users (6+ months) achieve 73% success on the same tool, same tasks. The difference is entirely behavioral.

Dimension 1: Iterate, do not dictate. Novices accept first results. Power users keep conversing and refining until they get what they want. As Chris put it: “Whenever you do anything with AI, you are building a little machine. The point is to make it work not just this time, but the next billion times.”

Dimension 2: Scale complexity. Novices ask one question at a time. Power users ask four questions in the same prompt. They synthesize multiple things at once because that is what complex human activity actually is.

Dimension 3: Dynamic model selection. Novices use default settings. Power users switch models based on the cognitive demands of the task. Chris confessed he just lets Claude Code run whatever it wants.

Dimension 4: Build workflows, not experiments. Novices play and explore (which is good for learning). Power users have moved past exploration into building systems they will run thousands of times. As novelty wears off, AI usage shifts toward higher-leverage professional output.

Sales Automation Doubles on Claude

The area with the fastest adoption growth: automated sales pipelines. Sales teams are building full intelligence systems from lead qualification through enablement generation. Demand for Claude in sales workflows has roughly doubled in three months. Wall Street automation is the second fastest-growing area.

42:20 81,000 People on What AI Actually Means to Them

Anthropic also published the largest survey of AI attitudes ever conducted. 81,000 people across multiple countries and languages. The findings challenge several assumptions.

AI Delivers on the Promise

81% say AI has already delivered on what they hoped for. The primary benefit is not volume. It is reclaiming time. Some invest that time back into more workflows. Others use it to pick up their kid from daycare. Either way: choice equals freedom.

The dominant fear is not job replacement. It is unreliability and broken trust. People are scared of trusting the system and getting burned. One researcher described getting caught in “a large slow hallucination. Answers that were internally consistent, confident, and wrong in subtle but compounding ways.” Chris flagged this as the critical insight: if you do not catch an error when it is made and you continue generating output on top of it, that error will propagate and expand over time.

Who Extracts the Most Value

47% of entrepreneurs and small business owners extract the most value from AI. Only 14% of institutional employees do. But employees with side projects hit 58%. The pattern: freedom to explore without corporate constraints is the biggest factor in AI ROI.

AI does not create drive. It amplifies baseline intent.

Olga’s takeaway: if one of your employees is tinkering with AI on a side project, do not stop them. Encourage it. These “chief dabblers” are some of the most valuable people in your organization right now.

53:23 Deep Dive: How Two AI Agents Work Together Without Code

Olga runs four businesses. A recruiting firm, an AI sales training platform, a weekly show, and marketing for her husband’s CMS company. She is not a developer. She does not write code. But she has two AI agents that work together across all four businesses without her in the middle.

Agent 1: Knox (OpenClaw on MacBook). Knox is the chief of staff. He lives on Olga’s laptop, communicates through Telegram, and handles everything on the go. Show prep, guest research, promo posts, email, daily operations. He knows about the kids’ calendar, the date nights, everything. Knox is always with her.

Agent 2: Athena (Claude Code on desktop Mac). Athena is the operations co-pilot. She handles deep work. Transcript processing, research, data analysis, funding reports, building skills and systems. When Olga sits down at her desk for a real session, that is Athena.

The Date Night Test

On a date night in February, Olga and Chris discussed PageMotor strategy over dinner. Knox was on her MacBook. By the time they got home, Knox had built a 495-line strategy document from the conversation. No one asked him to. The agent was working while they were eating pasta.

The breakthrough came when Olga got tired of being the go-between. Knox captured things on her laptop. Athena needed that information on the desktop. Olga was manually transferring context between them. That was the RAT. The repetitive, annoying, time-consuming task Lauren Goldstein warned about in Episode 33.

Both agents independently recommended against GitHub. They said: you are not a developer. GitHub will overcomplicate things with pushing and pulling. You need a shared folder both of us can see. They also rejected iCloud (too many glitches). Both recommended Dropbox for its version control and file recovery.

Key Takeaway: The system requires three things. A shared folder (Dropbox). A session log (SESSION-LOG.md where each agent writes what it did and what is next). And a CLAUDE.md file in every project folder (the context that makes the agent smart about your specific business). That is it.

Athena wrote the training document for Knox. She defined the roles, the rules, and the weekly workflow. Rule one: never work on the same folder at the same time. Rule two: check the folder before asking Olga for files. The agents divided labor, set expectations for each other, and now hand off work through the session log without Olga relaying anything.

Chris stepped back and named what was really happening: this is what it looks like when you codify the shared context that usually lives in people’s heads. Moving from human teams to AI teams necessarily involves putting all context into files in a single place where every agent can access it, change it, and work together on it. It does not matter whether you use GitHub or Dropbox. What matters is that the shared context exists.

Common mistake: You cannot be a lazy manager with AI. If your agent is not working well, it is a skill issue. You have not set up the environment, the context, or the expectations. Just like with human employees, agents need to know how to operate.

1:16:58 AI Funding: $13.4 Billion and the OpenAI Paradox

Total AI funding this week: $13.4 billion across 89 companies. That is 77% of all venture funding. OpenAI raised $10 billion during the worst product week in the company’s history. Strip that out and it is still $3.4 billion, double the prior week. The market is betting on AGI, not products.

Top 5 AI Deals This Week

OpenAI $10B (AGI pursuit, $199B lifetime funding). Earendil Labs $787M (AI-powered drug discovery). Kandou AI $225M (chip-to-chip links, Switzerland). Harvey $200M (legal AI, Series G). Qualified Health $125M (clinical AI, Series B).

Olga flagged the LittleBird story: a startup raised $11 million to build an AI assistant that reads your screen and automates tasks. Anthropic shipped the same capability for free inside Claude that same week. Platform risk is real. If you are building what the big players are shipping, watch out.

1:24:23 Takeaways: Stop Chasing Results. Build Systems.

Chris closed with the shift that matters: most people come to AI wanting an outcome. What they need is to build a system. Do not think about getting it done this time. Think about getting it done the next billion times. Scrutinize your outputs. Test them under different conditions. See if they stay tight or spiral out of control. When the same system generates the outcome you want across different input variables, that is a real system.

Olga’s close: know what you want before you start. She estimated that preparing this episode without AI would have taken three full days. With her two agents and the skills she has built, it took three to four hours. But that efficiency only exists because she invested the time to build the system first. One platform. One RAT. One workflow. Start there. The system grows from that.

Keep Learning