the bottleneck was never the coding

OpenAI posted about Symphony this week, and it has been stuck in my head.

The basic idea is simple: stop treating Codex like a thing you babysit in a terminal. Turn the project board into the control plane. Issues come in, agents pick them up, work happens in separate workspaces, and humans review the result.

That sounds obvious once someone says it.

It also feels like a pretty big shift.

I have been messing with a smaller version of that idea inside Transcripted. GitHub issues, labels, agent runners, review packets, human review. Nothing as clean as Symphony yet. More duct tape. More weird local scripts. More “why is this label stale” moments.

But the shape is starting to rhyme.

And it is making me think differently.

The bottleneck was never really the coding.

Or maybe it was, for a while. Coding used to be slow. I had to open the repo, find the files, make the change, run the tests, write the PR, and remember what I was even trying to do.

Now the agent can do a lot of that.

So the question changes.

Not: can the agent write code?

More like: can I make the work clear enough that an agent can pick it up without me hovering over it?

On April 29, I merged PRs #606, #608, #609, #611, #612, and #613, and still got 1.1.27 out the door.

That sounds like the good part.

And it is.

But the weird part is that the faster the coding got, the more obvious everything else became.

Not writing code.

Deciding what the work even is.

Is this issue clear enough for an agent?

Does it need a test?

Does it need a screenshot?

Is this actually two issues pretending to be one?

Can I review this in five minutes, or did I just create homework for myself?

That is the part Symphony helped name for me.

When everything was slower, messy thinking was easier to hide. A vague issue would sit there. A half-formed idea would just be a note. Bad backlog hygiene did not hurt as much because the whole thing moved slowly anyway.

But once the agent can turn a clean issue into a PR pretty quickly, the messy parts get louder.

The machine is not really asking me for better code.

It is asking me for clearer asks.

That is a different job than I thought I was signing up for.

That is why Symphony feels inspiring to me. Not because it means “AI replaces engineers” or whatever boring headline people will slap on it.

It is inspiring because it points at a different way to work.

Less session babysitting.

More issue shaping.

Less “please change this file.”

More “here is the outcome, here is how we know it is good, bring me something I can review.”

I wanted the breakthrough to be simple. The agents can code now. Cool.

Instead, the thing I keep learning is more awkward:

If I want better agent output, I have to become better at saying what I mean.

That is less flashy.

Probably more useful.