Zapier CEO Just Revealed What Most People Get Wrong About AI Agents

And how I turned my failed AI Agent into a workflow that actually works.

I spent three weeks building the perfect content agent.

It was supposed to handle everything: research my topics, extract key insights from my newsletters, transform them into social posts, and schedule everything automatically. Set it and forget it. The autonomous AI dream.

Well, I bet you can guess the result. It failed spectacularly.

Not because the technology wasn’t good enough. Not because I didn’t give it enough instructions. It failed because I was solving the wrong problem.

The agent would invent statistics that sounded credible but were completely wrong. It would mix up sources, blend different newsletter themes together, and publish content that sounded like me but felt... off.

And as you can imagine, this got worse. For every hour it “saved” me, I spent three hours cleaning up the mess.

Then I stumbled on a recent podcast by Peter Yang with Wade Foster, the CEO of Zapier, a company that’s been in the automation business longer than most of us have been using AI. He said something that stopped me cold:

“You probably don’t want an agent. You want a workflow that thinks.”

That one sentence hit me to the bone.

For someone who’s been obsessively following the AI agent industry, here’s what I’ve realized:

The entire AI community is obsessed with autonomy right now, agents that can “figure it out,” systems that make decisions without you, the complicated n8n workflows you see on YouTube, and the promise of complete automation where you just set a goal and walk away.

But that’s not what actually works in reality—at least not at the moment.

Because what I really needed was a workflow that follows the same proven process, executed consistently, with AI making smart decisions at specific moments.

I’m not saying my AI agent wasn’t smart enough to help me do that. It had tools, memory, and context. But still, the result didn’t actually hit me.

Sometimes the results were brilliant.

Sometimes they were garbage.

Always, they were unpredictable.

That’s when I stopped building an agent and started building a workflow that thinks.

Let me show you the difference.

The “God Agent” mistake

When I first started experimenting with autonomous AI agents, I had this vision: One prompt to rule them all.

I wanted an agent that could take my newsletter, understand my brand voice, research relevant LinkedIn trends, extract the key insights, format them correctly, and create social content that would perform well. Just give it the goal and let it handle the details.

The appeal was irresistible: Complete automation without having to think about process. Sounds cool, right?

So I built it. Spent hours crafting the perfect meta-prompt. Gave it access to my writing samples, my engagement metrics, my content guidelines. Fed it examples of high-performing posts.

And guess what? It didn’t work, at least not at the level of quality I wanted:

1. The hook was weak even though the insight was on‑brand, and the format was too long and redundant.

2. It got worse when it invented a statistic about “73% of knowledge workers” doing something. It sounded credible, and was completely made up.

3. Then it mixed up two different newsletter topics and created a Frankenstein post that made no sense.

Almost every run was different. I had no idea what I’d get until I reviewed the output. And reviewing took almost as long as just writing the post myself.

---

A Quick Note: I’m part of Cozora, where AI builders share real implementations and frameworks that actually work. If you’re serious about mastering AI like the ones in this post, AI Maker Lab members get 50% off annual membership.

---

Why even sophisticated AI agents fail without workflow structure

Now, you might be wondering: why is this happening?

The reason is actually simple. It’s because the AI agent is making complex judgment calls without the context to make them well.

Let me explain.

1. Autonomy compounds judgment errors

Creating content isn’t a task with objectively right answers. Every decision has its nuance:

• Which insight will resonate with my audience? (Multiple valid options)

• What makes this research source “credible enough”? (Subjective quality bar)

• Does this hook create the right KIND of tension? (Strategic positioning)

When an agent makes these calls autonomously, it’s guessing at context it doesn’t have. Research the topic? It might find credible sources that don’t align with my audience’s beliefs. Extract insights? It might choose one that’s factually correct but strategically wrong for my positioning.

Each autonomous judgment multiplies the chance of subtle drift. And you won’t know until you review the final output—by which point all those judgment calls are baked in.

2. No visibility into the judgment calls that matter

When my agent produced generic content, I couldn’t tell which decision went wrong:

• Did it choose the wrong insight in the first place?

• Did it misinterpret what makes my voice unique?

• Did it misjudge what level of research depth my audience expects?

I just had mediocre output and no clear path to fix it. With complex work, you can’t just regenerate and hope for better luck. You need to know WHICH judgment went wrong and WHY.

3. Quality control after all decisions are made

The agent ran from start to finish without stopping. By the time I could review anything, it had already made 20+ judgment calls about nuance, positioning, and strategic intent.

If the hook was weak, I couldn’t adjust just that—the entire post was built on the insight the agent chose in step 1. If the research was shallow, I couldn’t add depth without regenerating everything.

For complex work, you need checkpoints at the moments where judgment actually matters.

The Agentic Workflow Framework That Changed Everything

Then I saw this spectrum from Wade Foster’s team at Zapier that made everything clear.

And before you ask: this framework works whether you’re building agent in Make, n8n, Zapier, or custom code. The principles are platform-agnostic:

Look at this carefully. There are four approaches:

• Far left: Traditional workflow automation with no AI. Fixed steps, zero intelligence.

• Left-center: AI Workflow. AI makes ONE smart decision within an otherwise deterministic path.

• Right-center: “Agentic” Workflow. AI makes MULTIPLE decisions, but still follows a defined path with clear steps.

• Far right: Agent. AI has tools, memory, knowledge, and decides its own path to accomplish a goal.

Here’s what I realized staring at this:

I had built a far-right agent. But I was asking it to figure out EVERYTHING on its own.

No wonder the result was a disaster.

Instead of asking, “create a LinkedIn post from my newsletter,” I should have defined a workflow like this:

Extract insight using this criteria → Select from these three patterns → Generate hook with this structure → Format using this template

Same AI capabilities. Same access to tools and context. But structured path instead of open-ended autonomy.

Where your AI Automation belongs on the spectrum

Here’s what the spectrum actually tells you: the complexity and nuance of your work determines where it belongs.

AI Workflow (left-center) - Simple judgment, clear execution:

• Work with ONE subjective decision point

• Example: New lead arrives → AI writes personalized message → Send SMS → Add to CRM

• The judgment: How to personalize this message based on lead data

• Why this works: The personalization decision is isolated; everything else is mechanical

Agentic Workflow (right-center) - Complex work with known process:

• Work with MULTIPLE judgment calls but a proven sequence

• Example: My newsletter workflow—extract insight, research data, synthesize argument, generate hook, build narrative, match conclusion

• The judgments: Which insight resonates? What’s credible? What creates tension? What conclusion type fits?

• Why this works: I know the steps that produce quality; AI makes smart decisions within that structure

Full Agent (far right) - Truly exploratory work:

• Work where you genuinely don’t know the path ahead of time

• Example: “Research this emerging technology and determine if it’s relevant to my business”

• The decisions: What to investigate, how deep to go, what questions to ask, what format makes sense

• Why it’s rare: Most knowledge work follows patterns you already know

But here’s what you need to keep in mind: For complex, nuanced work, skip full autonomy. Use structure, and make room for sharp judgment at the right checkpoints.

AI Agent vs Agentic Workflow: What This Looks Like in Practice

Let me show you what I actually built to replace my failed agent. This is my newsletter-to-social workflow—the one that produces consistent, high-quality content.

The Agent Approach (What I Was Doing)

Single goal: “Create LinkedIn content from my newsletter that will perform well.”

What the agent had access to:

• My complete content library

• Brand voice guidelines

• Engagement metrics from past posts

• Examples of high-performing content

• Tools to research trends

What the agent decided autonomously:

• Whether to research current trends first

• What to extract from the newsletter

• How to format it

• What hook style to use

• What length to target

• What call-to-action to include

• Whether to reference other content

Result: Unpredictable quality, inconsistent voice, invented facts, missed outputs.

Well, it’s different every time.

The Agentic Workflow Approach (What I Built Instead)

This is a fixed path with multiple AI decision points. And this is what I’ve been sharing on Maker Labs:

Step 1: Content Analysis → Extract Main Contrarian Insights

• Structure: Pull newsletter content, parse into semantic sections (intro, frameworks, examples, conclusions)

• AI Decision: “Analyze sections 2-4. Extract 2-3 contrarian insights that challenge conventional wisdom about AI, productivity, or work. Each insight must directly contradict a common belief. Rank by controversy potential.”

• Guardrails: Must identify contradictions, must be from specified sections, must rank by impact

• Output: Ranked list of 2-3 contrarian insights with reasoning

Step 2: Research & Validation → Find Supporting Data

• Structure: Take top-ranked insight, trigger web research across news, industry reports, competitor analysis

• AI Decision: “Search for: (1) Recent data/statistics that support this insight, (2) Examples of companies/people proving this point, (3) Counterarguments to address. Prioritize credible sources from last 6 months.”

• Guardrails: Must find quantitative data, must cite sources, must include counterarguments

• Output: Research summary with citations, data points, and opposing views

Step 3: Argument Synthesis → Strengthen The Case

• Structure: Combine original insight with research findings

• AI Decision: “Integrate research data into the original insight. Add specific examples or statistics that make the claim more concrete and credible. Identify the strongest supporting evidence and the most important counterargument to address.”

• Guardrails: Must maintain original insight, must cite specific data, must acknowledge limitations

• Output: Strengthened argument with embedded evidence

Step 4: Hook Generation → Create Bold Statement

• Structure: Analyze argument structure (reversal, comparison, prediction, contradiction)

• AI Decision: “Generate a bold statement hook that creates immediate tension. Use concrete data from research if available. Make it scroll-stopping by creating cognitive dissonance. Maximum 2 sentences.”

• Guardrails: Must use research data if available, must create tension, 2 sentence maximum

• Output: Hook with optional data point

Step 5: Post Structure → Build Narrative Arc

• Structure: Define post sections (hook → setup → framework → evidence → conclusion)

• AI Decision: “Create 3-4 body sections that: (1) Explain why conventional wisdom exists, (2) Present the contrarian insight with evidence, (3) Show practical implications, (4) Address main counterargument. Each section 2-3 sentences maximum.”

• Guardrails: Must follow narrative arc, must address counterargument, section length limits

• Output: Structured body sections with flow

Step 6: Conclusion Matching → Content-Appropriate Takeaway

• Structure: Analyze post category (business strategy, workflow, thinking framework)

• AI Decision: “Determine post type. If business/strategy → strategic insight conclusion. If workflow/tool → implementation framework. If thinking → cognitive shift. Write conclusion that matches content type and delivers THIS story’s specific insight.”

• Guardrails: Must match content category, must avoid generic advice, must deliver specific takeaway

• Output: Category-matched conclusion

Step 7: Voice Calibration → Ensure Brand Alignment

• Structure: Load brand voice samples and anti-patterns

• AI Decision: “Review complete post against voice guidelines. Check for: (1) Contrasting ‘not X, but Y’ patterns (remove), (2) Corporate language (simplify), (3) Generic productivity advice (make specific). Flag and fix violations.”

• Guardrails: Must remove anti-patterns, must maintain conversational tone

• Output: Voice-calibrated post

Step 8: Human Review Gate

• Structure: Send to review queue with all components visible (insight → research → hook → body → conclusion)

• Human Decision: Approve, edit specific steps (regenerate research, adjust hook, modify conclusion), or restart from different insight

• Output: Published or edited post

Why this works better

Here are three reasons why agentic workflow is better than agent:

1. Reliability through structure

Same 8 steps every time. Research always happens (Step 2), hook always uses that research (Step 4), conclusion always matches content type (Step 6). I know exactly what I’ll get—not which format, but which process. The variation happens where I want it: the specific insights, the data found, the synthesis quality.

2. Debuggable at every stage

When something’s off, I know exactly which step failed:

• “Step 2 research didn’t find strong enough data” → Refine search parameters

• “Step 4 hook lacks tension despite good research” → Adjust hook generation guardrails

• “Step 6 conclusion defaulted to generic advice” → Improve content category detection

By having a step-by-step process, everything becomes fixable.

3. Quality control gates at critical points

I can review AI decisions at eight specific moments in the workflow. If something’s off at any checkpoint, I regenerate just that step with different parameters. Everything else stays intact.

Why complex AI Automation can’t be fully autonomous

Here’s the part most people miss when they chase autonomous agents:

“Not all work has the same structure.”

Simple tasks with clear right/wrong answers? AI can handle those autonomously. Data entry. Email categorization. Format conversion.

But complex knowledge work—the kind that actually creates value—has more complexity than we even realize:

1. Multiple valid paths, no “correct” answer

When a lawyer creates legal strategy, there are five defensible approaches. When a designer chooses a direction, three concepts could all work. When a strategist positions a product, multiple angles are valid.

But the real job isn’t about finding the right answer. It’s about judgment:

“Which approach fits this client’s risk tolerance? This brand’s identity? This market moment?”

AI can generate options. Only humans can make the judgment call about which option serves the strategic intent.

2. Quality depends on context AI doesn’t possess

What makes communication “persuasive” varies by audience. What makes research “credible” depends on industry norms. What makes a decision “bold vs. reckless” depends on company culture.

These are all derived from contextual judgments based on an understanding of people, culture, and politics that AI can’t access from task data alone.

3. Strategic intent shapes everything

The same input produces different outputs depending on invisible intent:

• Writing to build authority needs different elements compare with writing to drive conversion.

• Designing for innovation requires different tradeoffs than designing for trust.

• Negotiating for long-term relationship has different tactics compared to closing a one-time deal.

AI can’t infer intent from the task itself. Strategic context lives in your head, shaped by goals AI can’t see.

Your move: Audit one AI interaction this week

Here’s what I want you to do, and I mean actually do, not just think about doing.

Pick one task where you’re currently using AI. Maybe it’s content creation, maybe it’s research, maybe it’s data analysis. Something you do regularly enough that getting it right matters.

Now map it against the spectrum by asking yourself these three questions:

1. Where does this task actually sit on the spectrum?

• Are you using a single AI prompt?

• Are you using an agent with tools and memory?

• Or something in between?

2. Where SHOULD this task sit based on its complexity?

• Simple judgment, clear execution? → AI Workflow works fine

• Multiple judgment calls, proven process you already know? → You need an agentic workflow

• Genuinely exploratory work with unknown path? → Full agent makes sense (but this is rarer than you think)

3. What’s the gap between where you are and where you should be?

• Using a full agent for work that needs structure? You’re getting unpredictable results

• Using simple prompts for complex work? You’re leaving capability on the table

Here’s your starting framework

If you identify work that needs an agentic workflow (and most complex knowledge work does), here’s how to build one:

Step 1: Map your current process

Write down the steps you ALREADY follow when this work goes well. You’re not inventing a new process—you’re documenting the one that’s already in your head.

For my newsletter workflow, I already knew my best content came from: extracting contrarian insights → researching supporting data → building narrative → matching the right conclusion type. I just made that explicit.

Step 2: Identify the AI decision points

Look at each step and ask: “What judgment call happens here?”

Those judgment points are where AI makes smart decisions within your structure:

• “Extract contrarian insights that challenge X”

• “Find recent data supporting Y”

• “Generate hook creating Z type of tension”

Step 3: Add guardrails at each decision point

Define what “good” looks like for each AI decision:

• Must cite sources

• Must use data from last 6 months

• Must create tension

Guardrails are important. They represents clarity. They tell AI exactly what judgment you want it to make.

Step 4: Build review gates at critical moments

You don’t need to review every step. But identify 2-3 moments where if the AI judgment is off, everything downstream breaks.

For me: If Step 2 research is weak, the whole post suffers. So I review research quality before moving to hook generation.

Step 5: Run it, refine it, evolve it

Your first version won’t be perfect. Mine wasn’t. But here’s the magic: Every time you refine one step, every future run gets better.

I’ve refined my workflow eight times. Step 2 now prioritizes quantitative data over anecdotes. Step 6 matches conclusions to content type instead of defaulting to generic advice. Each improvement compounds.

Start small, think structure

So here’s my question for you:

What’s the one AI task you’re currently treating like a magic box, hoping for good results but not sure why they vary that you could rebuild as a structured workflow this week?

Don’t try to rebuild everything at once. Pick ONE task where you’re currently getting inconsistent AI results.

Map the process you already know works. Add AI decision points with clear guardrails. Build review gates at moments that matter.

Run it five times. You’ll spot exactly where to refine.

That’s how you build AI systems that get better instead of just different each time.

Next week, I’m gonna show you how to build an AI agent for your personal productivity.

Stay tuned!