How do AI coding agents work?
🧠 The core technologies involved, how they work and exploring the risks of AI feature slop. Knowledge Series #74
🔒The Knowledge Series is available for paid subscribers. Get full ongoing access to 70+ explainers and AI tutorials to grow your technical knowledge at work. New guides added every month.
Since the arrival of Claude 4, some engineers have boasted that they’ve been able to clear their entire backlogs for the first time:
In the future, the concept of a product backlog may not exist as we know it today. Why would you need a backlog if your biggest constraint (engineering resource) is 50x’d by AI?
AI coding agents could eventually make this possible and GitHub recently joined the growing number of companies with their own fully fledged AI coding agents that engineers can assign work to directly.
This could have all kinds of consequences. Resource constraints are often used by product teams to manage the expectations about what can be built. If this is removed completely, there’s a real risk of AI feature bloat, or AI-powered “feature slop”. On the plus side, when conversations about resource constraints become less important, the focus may shift away from what’s feasible based on resources and more sharply towards what’s actually valuable for users.
In this Knowledge Series, we’re going to take a closer look at how these AI coding agents actually work. We’ll explore the core technologies behind them, the key aspects of the software engineering SWE-bench test and how assigning Jira / Linear tasks to an autonomous agent actually works with real world examples.
Coming up:
What are AI coding agents?
The core technologies used explained: LLMs, MCP, code repos and more
The Software Engineering SWE-bench test explained: a closer look at how AI engineering agent capabilities are assessed
Real world examples of how companies and engineers are using AI coding agents
35+ companies actively using AI during their engineering process
The future of AI coding agents: opportunities and risks for product teams
What are AI coding agents and how do they work?
AI coding tools like GitHub’s Copilot have been around for a few years now but AI coding agents take these capabilities to the next level. Rather than merely suggesting code or generating code inside applications with AI, coding agents actually allow engineers to assign specific tasks to them for completion.
The tech stack of a coding agent will vary depending on the tool a product team is using but here’s a snapshot of what a typical AI coding agent stack might look like:
The process starts with an engineer assigning work to an AI agent in their integrated development environment (IDE) or directly from GitHub (through Github.com ) . Once they’ve been assigned a task to work on, a coding agent will gather the necessary requirements, analyze codebases, develop a plan of action, write the code in a secure environment and ultimately generate a pull request for review by a human engineer.
The types of tasks that are currently given to AI Agents are currently being framed by many companies as the “things that engineers don’t want to do” - think bug fixes, test coverage, refactoring, API endpoint creation etc. But in the future, the remit of the kinds of work AI coding agents can do will inevitably expand.
The core technologies used - a simple explanation of the important bits worth knowing
Before we take a closer look at the test that’s used to assess these agents along with some real world examples, let’s break down the core technologies used in a typical AI coding agent set up: