With vibe coding, someone might ask, "create a sales website," and it might work, but not have security, stability, etc. issues.
With context engineering, developers first think through the approach with AI (including security, stability, scaling, etc), get the agent to document what they discussed, then close the agent, and open a fresh one, to clear out the old memories.
This is the key. They manage the context (info in the agent's memory) to keep it from filling up and containing different ideas. This limits hallucinations and keeps them focused on one task.
Then, they ask the fresh one to use the design to come up with the plan. After the detailed plan, with check boxes, is developed, they have it written out, and then shut the agent.
Then, they run a loop where a new agent is spawned, works on one item, checks it off, creates notes for its successor, then shuts it down, and spawns a fresh one (with no memory) to read the notes and take off from where the previous one left off. It's thus a fully automated, but controlled process.
© 2025 Praveen Puri
