A major drawback of the Agent is that it just writes the code without consulting the developer first or doing any planning. This is a waste of time and credits because often such workflow will produce unexplainable or simply bad code. Adding to that is the fact that Claude is extremely overzealous with assumptions.
This results in a workflow that’s nothing like working in a team with other developers. Most humans’ workflow includes the following steps (in order):
- Task definition
- Research
- Planning implementation strategy
- Implementing
- Testing
Agent mode only implements while missing context from the remaining steps.
Solution: A combination of modes in 3 steps
- Planning. Agent finds and reads relevant files in the codebase autonomously and proposes a solution “on paper”.
- Approval (inspired by Replit Agent). The agent lays out a plan with tasks listed as optional checkboxes. The user can either continue the discussion or select from a list of relevant tasks. Finally, the user approves the plan.
- Implementation - Existing Agent mode.
This would consume the same or fewer amount of credits, while providing a much higher quality experience.
Instead of correcting mistakes and countless assumptions made by the agent. The Planner allows developers to take control of the project and have an AI-powered workflow more akin to working with a qualified junior developer, rather than a clueless child.
What’s required for this to work:
- Agent asks clarifying questions
- Agent researches and demands additional information where assumptions are likely to be made otherwise
- Agent requires explicit approval before making changes
- Agent follows the instructions built up in context until this point while implementing