The 12-Line Pull Request That Took 5 Days: A Context Problem

Dennis Pilarinos

November 25, 2024

TL;DR: The biggest challenge in software development isn’t writing code. It’s having enough context to know what code to write

The state of software development is still unproductive

Every year Stack Overflow releases their developer survey, and year after year the results remain the same: most developers feel that they are not as productive as they wish they could be.

This year, 81% of the 65,437 developers surveyed said that productivity is the number one area where they want to see an impact from AI.

Yet, in a year that has seen an explosion in AI dev tools, Stack Overflow found that:

Siloed information still prevents developers from shipping code
- 53% of developers are blocked by knowledge gaps
Developers continue to waste too much time searching for the answers they need to do their jobs
- 63% spend more than 30 minutes a day looking for information
Answering questions from others on a team is an ongoing source of distractions
- ‍49% of developers lose 30-120+ minutes a day giving others information about how their codebase works

This data raises the question: for all of AI’s claims about increasing developer productivity, why aren't developers seeing more of an impact?

Software’s biggest challenge is context, not code

In some ways, the gap in expectations versus reality around AI that the Stack Overflow survey highlights is not surprising.

Companies invest a great deal in hiring engineers who are skilled at writing production-level code. Even so, the vast majority of AI developer tools on the market today are focused on code generation.

There are lots of tools to help us write code, but what we need are tools that give us the answers we need, so we know what code to write.

For most teams, the process for building software can be split into three iterative stages:

Gathering Context
- Developers seek to answer questions like "How does the system currently function? What was the original intent behind certain design choices? What are the potential ripple effects of making a change?"
Implementing Solutions
- This is where the process of translating context into code happens. Code generation tools are focused on this stage and produce code suggestions
Operating Services
- As the problem is understood and the solution is codified, teams typically deploy, operate, and monitor their service at this stage

One of my favorite anecdotes that highlights this point is from an engineering leader at a company known for hiring very competent developers.

"It took 5 days to write a 12 line pull-request. They didn't need help with writing code. They spent the week gathering the context, history, and information about their application to be able to write those 12 lines."

Today’s code generation tools aren’t built to address this particular challenge.

If we want to see actual productivity benefits from AI tools, we need to direct this technology toward our biggest problems.

Our codebases are a compilation of thousands of decisions, discussions, and documents that live across tools like Confluence, Slack, Notion, Jira, Linear, GitHub, and more.

Before we write code, we need the context that these tools contain - this is where AI can make the biggest difference.

From search engines to context engines

Every tool we use to develop our application contains a slice of the context that makes up the whole picture of how and why things work the way they do. Without access to this context, individual developers are like the characters in the story of the blind men and the elephant - all acting on incomplete information.

Search doesn’t offer a much better alternative because it's too time consuming and error prone. Plus, developers need to know where to look and what exactly to look for. New team members are especially challenged when asked to rely on search to understand a codebase. If you weren’t there when the original decision took place, it’s unlikely that you will know what context exists, let alone where that context lives.

This is why we built Unblocked - because we believe we can make developers productive by providing helpful answers to their questions about their codebase.

Unblocked augments your source code with relevant knowledge across your internal systems (GitHub, Slack, Confluence, Jira, Notion, and more) all while respecting the underlying permissions models. No matter the question, we can then explain the nuances of your codebase — how it works, why it was written, and why it works the way it does.

The impact here is profound: with Unblocked, developers can answer their own questions without interrupting their peers or having to wait for meetings. What’s more, they can find the entirety of the information they need wherever they're working (including in their IDE).

Because of this, our customers tell us that Unblocked saves an hour or more of time per day for every engineer on the team.

Toward a future where time to knowledge is near immediate

The best developer tools will be the ones that give developers superpowers to do their work.

Many AI tools will continue to assist with the process of writing code, but regardless of how far AI code generation tools go, developers will remain an integral part of the software development process.

Unblocked's goal is to help developers quickly get the information they need so they spend less time digging for information or interrupting their coworkers and more time writing code.

If you haven’t started using Unblocked, book a demo with our team.

‍