@cloneisjun I think they solve different problems.
Instructions can be more cost-efficient for simple tasks. But for complex work, a well-designed graph can be a game changer.
To me, it’s a new perspective in the evolution of AI engineering from telling AI what to do > designing how it works
We went from prompt engineering → context engineering → loops.
Now everyone is talking about graph engineering.
But I think there's a more useful way to understand what's actually changing.
🧵
AI engineering is starting to look less like prompt writing...
…and more like workflow architecture.
The interesting question isn't just:
"What can my agent do?"
It's:
"How should the work flow through my agents?"
So maybe the real skill isn't building the most complicated agent graph.
It's knowing when you don't need one.
A simple workflow that reliably ships software beats a beautiful architecture nobody understands.
Simplicity still wins.
But there's a trap.
More agents ≠ better system.
Every additional agent can add:
• context
• tokens
• cost
• coordination
• failure modes
You can easily build an incredibly sophisticated system...
that is worse than a simple one.
So we started giving agents context.
Project rules.
Architecture decisions.
Documentation.
Design requirements.
Files like CLAUDE.md, AGENTS.md, and project specs.
The agent wasn't just getting a prompt anymore.
It was getting a working environment.
working at a big company is basically an accelerated course in how power actually works. once you see the sheer level of inefficiency, rent-seeking, & arbitrary decision-making, it kind of breaks all illusions about big companies.
the best part is realizing that half the people in charge have no clue what they’re doing but just sound confident. once you internalize that, you stop overestimating the competition & start realizing that most barriers to entry are just psychological.