DEV Community

Dor Amir
Dor Amir

Posted on

Draft: Why Your AI Agents Need a Tech Lead

#webdev #ai #programming #productivity

You've got Cursor. You've got Claude Code. Maybe you've got a couple of aider instances running. Your AI coding setup is serious.

And somehow, shipping still takes forever.

The tools aren't the problem. The coordination is.

The missing layer

Here's what a good tech lead actually does, stripped of the politics:

  1. Reads the codebase and understands what's there
  2. Takes a vague feature request and turns it into specific, scoped tickets
  3. Figures out which engineer (or agent) should handle what
  4. Reviews the diff when it comes back, catches the stuff that works technically but breaks the system conceptually
  5. Merges or sends back with comments

AI agents are excellent at step 4 when the ticket is good. They're bad at steps 1, 2, and 3. And they can't do step 5 at all, because they don't know what they didn't write.

You're doing steps 1-3 manually every time you prompt. That's the bottleneck.

What Draft does

Draft is an AI tech lead. You describe a feature in plain English. It reads your codebase, generates the tickets, delegates to AI agents, and reviews the diffs before they land.

You: "Add rate limiting to the API endpoints"
Draft: [reads your codebase]
       [generates 3 tickets: middleware setup, config, tests]
       [delegates each to an agent]
       [reviews diffs for consistency with your existing patterns]
       [surfaces the one that introduces a breaking change]
Enter fullscreen mode Exit fullscreen mode

You stay in the loop where it matters. The mechanical overhead disappears.

Stack is FastAPI + React + SQLite. Self-hostable. No black box SaaS absorbing your codebase.

Why this matters now

LLM coding quality improved fast. Context windows got huge. The agents got good.

The problem shifted from "can the AI write code" to "can the AI write the right code, in the right place, without breaking what's already there."

That's not an LLM problem. That's an architecture problem. It's about having a clear picture of the system before you start, and a skeptical eye when the diff comes back.

A tech lead isn't valuable because they can write code faster than you. They're valuable because they've read everything, they know what the code is supposed to do, and they can tell when something is wrong even if it compiles fine.

Draft is that layer for your AI stack.

The alternative

The alternative is you. You read the codebase before every session. You write the prompts. You review the diffs. You catch the regressions.

That works. It's just slow, and it doesn't scale past what one person can hold in their head.

If you're running multiple agents in parallel, you need something coordinating them. Otherwise you end up with three agents that each "fixed" the same function in different directions and now you have a merge conflict and no one to blame.

Draft is at trydraft.dev. (Full disclosure: I work on the team.)

Try describing a feature and see what it generates for your codebase. The ticket breakdown alone is usually worth it.

Top comments (0)