Let’s talk about agentic workflows for a minute.
One of the biggest mistakes people are making right now is building closed loops into their agent systems. OpenManus is a great example of how not to do this.
Here’s what’s going wrong:
🔹 Vague prompting.
Telling an AI “you’re an autonomous programmer” is like saying, “you’re a doctor—now go do brain surgery.”
Without specificity, you're not getting predictable outputs.
If you're asking an AI to reason through code generation, define what “good code” looks like.
What make-or-break patterns does your project need to succeed? Need SOLID/DI/MVC and Singletons? Great—say that in the system prompt.
Set expectations for documentation, commenting, modularity, and error handling.
Assume you're working with a perfect stranger—because you are.
Pro tip: Embedding literary influences that deal in recursion concepts can help avoid broken debugging loops. Not kidding.
🔹 Single-task agents that don’t communicate.
I’m not actually against focused, single-task agents.
But if you need multiple concurrent task focused agents, then you also need team dynamics.
Organized teams are focused communication machines.
But here?
You’ve got a planner that can’t code. A browser that can’t plan. A coder that can’t browse.
And none of them talk to each other.
Why design it this way?
Just like in real life, siloed teams lead to incoherence, dead ends, and loops that burn cycles without producing output.
If your agent can execute a 20-step process, then it can also vary its system prompt and temperature per step without losing your rich context.
This isn’t hard.
If every step is handled by a rigid agent without context or adaptability, you're going to hit entire categories of tasks where they can’t reason through the problem space.
Here’s the fix:
Instead of:
Step 1: <coding agent implements>
Step 2: <coding agent debugs>
Step 3: Repeat Step 1, hoping something works by Step 20...
You want:
Step 1: <agent implements>
Step 2: <agent debugs>
Step 3: <alternate agent observes + interjects with context>
“Looks like you're stuck using npx. Want to try npm instead?”
Stop thinking about speed per task.
Start thinking about forward momentum and process integrity.
Open loops that allow variation, feedback, and contextual adaptation will always outperform rigid pipelines.
Sure, a task might take an hour.
But a well-structured loop finishes.
A bad one eats tokens in a recursive death spiral.
Need help designing AI systems and agentic workflows that don’t self-destruct?
I build them every day.
What kind of crazy awesome AI projects are you imagining?