"Plan and build a dashboard for me." Sounds like it should work. It doesn't.
Planners and developers have different goals.
A planner's goal is delivering value. "What will the customer be able to decide from this?"
A developer's goal is shipping code. "Does it run?"
Give one agent both jobs, and it prioritizes whichever is easier to measure.
"Is the planning good?" Hard to judge.
"Does the code run?" Immediate feedback.
So the agent rushes through planning and jumps straight to implementation.
"Plan and build this" becomes "half-ass the plan and build this."
That's why role separation matters.
1️⃣ Researcher role → Investigate customer interests, gather industry trends and benchmarks
2️⃣ Planner role → Redefine customer value based on research, select key KPIs, design screen structure
3️⃣ Developer role → Implement based on the planning doc
We review each agent's output, iterate with feedback and direction until it hits the right level, then move to the next stage.
Now each agent can focus on its specialty.
Once the planning agent finishes the "why," the dev agent only worries about the "how."
Separate the roles. Define each role's expertise and success criteria.
That's context engineering.
What's interesting is how much context engineering resembles leadership.
A leader isn't a superhero who's best at everything.
A leader sets a valuable goal and assembles a team of specialists who are better than them.
AI Agents aren't genies—but with a good leader, they become an incredible team.
Top comments (0)