Skip to content

idrankvodka

Thoughts, after a sip or two.

The AI Circle

Agents Reviewing Agents

Honest Dev say: "The robots are working. I am watching the robots work."

In my last post, I talked about how AI is here whether people like it or not. I still believe that. But the way I have seen it being used since then has been... something.

The Setup (A True Story)

I am currently on a contract with two senior developers. These are not juniors figuring things out. These are seasoned engineers. They have opinions about architecture. They have seen things.

They also brag about how many hours they spend setting up their AI coding workflows.

Not building features. Setting up workflows.

  • The Stack
    • Linear for project management (which has an MCP server).
    • Editor connected to Linear via MCP.
    • Write tickets in the editor. Via MCP.
    • Pull tickets back into the editor. Via MCP.
    • Set off AI agents to write code. Multiple branches. Worktrees.
    • An architect agent. A code reviewer agent. A security expert agent. etc.

It sounds impressive when you list it out like that.

The Daily Routine

Here is what actually happens every day:

  1. AI agent writes the code.
  2. They sit and monitor the AI agent writing the code.
  3. Code is done. They run the review agent.
  4. Review agent spits out feedback.
  5. They apply the fixes.
  6. Push to GitHub.
  7. Request a Copilot review on the PR.
  8. Copilot spits out its own feedback.
  9. They pull Copilot's feedback back into the editor.
  10. They have their agents re-review Copilot's suggestions.
  11. Apply more changes.
  12. Push again.
  13. Repeat.

This is the entire day. Every day.

To summarize: An AI writes code. An AI reviews the code. Another AI reviews the review. A human monitors all of this. The human is called a "senior developer."

The Cost

These guys have at least four AI subscriptions each. Mid to top-tier plans. We are talking roughly $700 USD per month. Per person.

  • What $700/month buys you
    • The privilege of watching agents work.
    • The ability to have AI review AI-generated code.
    • The comfort of knowing your workflow is "optimized."
    • A very expensive loop.

The "Optimal Setup" Argument

I know what some people will say. "Well, you need to have an optimal setup." Fair enough. But what is optimal here?

  • I have watched people spend the time they could use to build an actual feature on implementing AI workflows instead.
  • Those workflows become obsolete the moment these AI companies release a new feature or deprecate an old one.
  • Then they rebuild the workflow.
  • Then they optimize the rebuilt workflow.
  • Then the next update drops.

It is a setup for setting up setups.

The Efficiency Question

You would think with all this power, with all these agents, with all these subscriptions, they would be shipping faster than everyone else.

Nope.

Just an endless circle. Agent writes. Agent reviews. Agent reviews the review. Human watches.

This is not what I had in mind when I wrote the other post. I said AI is a tool. I still think it is a tool. But I did not expect people to build Rube Goldberg machines with it.

Non-Conclusion (Again)

I have no conclusion. I have no advice.

It is funny. It is sad.

Is this what you guys are experiencing too?

Still observing. Still concerned. (Still building with whatever tools work).