How to Use LLM Agents Without Losing Your Mind (Or Your Dignity)

Software is all leverage. LLM agents amplify leverage. Here is how I accelerate my development.

The Four-Phase Approach (Because Five Would Be Showing Off)

1. Research
2. Plan
3. Execute
4. Test

Phase 1: Research

• History: Why does this exist?
• Know what you’re actually solving.
• Understand the architecture.
• Gather your artifacts: error logs, stack traces, code snippets, API docs, schemas.

Phase 2: Plan

• Architect the design changes first.
• Specify requirements.
• Define the boundaries.
• Split the plan so that smaller agents can concurrently implement the feature.

Phase 3: Execute

• Reference the plan repeatedly.
• Context is the parameter you are optimizing. For the agent implementing database changes, it just needs the database architecture and the boundary.
• Get a separate agent to implement the tests.

Phase 4: Test

• Read the code. If it is going to production, your eyeballs need to review it.
• Does it do what you think it does?

What Not to Do: A Cautionary Tale

• Prompt Roulette: We are engineering, not gambling.
• “Please Fix”: It will band-aid, not RCA.

The dirty secret nobody wants to admit: LLM agents are tools, not teammates. Powerful tools, sure. But tools nonetheless. Master the process, respect the limitations, and maybe—just maybe—you’ll ship something that doesn’t make you hate yourself in the morning.

Now get back to work.