The last mile is reading
For most of the history of this job, typing the code was the expensive step. Tools were built around it: completion, snippets, refactorings, generators. Review was the cheap step at the end, a colleague glancing over a few hundred lines they could have written themselves.
Coding agents turned that around. We use them every day and we are not going back. An agent given a clear task writes a working feature while you get coffee. It also writes eight hundred lines to do it, touches fourteen files, and is just as confident about the part it got wrong as about the part it got right.
The bottleneck moved
Somebody still has to read it. Not because agents are bad, but because somebody answers for the code that ships, and that somebody is a person. In practice it is two people: the developer who asked for the change, and the lead who approves the pull request.
Both of them are now reading far more code than they used to, with tools designed for far less. A pull request page is a list of files in alphabetical order. It tells you nothing about which three files carry the change and which eleven are fallout. It cannot answer “who else calls this” without a checkout. Its discussions scatter across tabs and notification mail.
So people skim. A review of an agent’s change becomes a scroll and an approval, and the one thing the human was there for does not happen.
What reading needs
We think a tool for this has to do a few things that a diff viewer does not:
- Show the shape first. Which types changed, what was added to them, what calls what. Ten seconds with a graph tells you where the next ten minutes should go.
- Take the noise away. A renamed parameter rippling through twenty files is one decision, not twenty. Files with no change in behaviour should be ticked off before you get there.
- Keep the code navigable. Go to declaration and find usages belong in a review as much as in an editor.
- Keep the conversation in one place. Every thread you are part of, with what is new since you last looked.
- Start earlier. The cheapest review is the one the author does on their own machine, before anyone else spends time on it, with the comments going straight back to the agent.
Why a whole platform
We started with a pull request review tool. It kept growing backwards: to review the agent’s output locally you need the working tree, an editor, the language servers; to give the agent a better task you need somewhere to write the task down and draw what you mean. That is how Deppo became a development platform. It covers planning, design, code and review in one native window.
The weight is still on the end. Everything before it exists so that the reading goes well.
Deppo is in closed beta. If this is the problem you have, ask for access.