Hello and welcome to this edition of Ctrl+Alt+Deploy! I’m Lauro Müller, and I’m super happy to have you here with me 😄 Let’s spend the next few minutes sharpening our skills and learning something new and relevant for running production systems 💪 Today we’re going to spend some time to dive deeper into how we can manage long-running coding sessions with agents more effectively. I hope you enjoy, and let’s get started!
If you’re working regularly with coding agents, you know they get worse over a long session: the model asks about a decision you settled an hour ago, reverts a fix it made twenty minutes ago, or writes something in a style you have already corrected twice. Or maybe that happens only with me? 😅
Most people respond by feeling, clearing when it gets bad enough, compacting when that little circle turns red, or delegating to a subagent because delegation is supposed to solve all the world's problems.
Those moves are not interchangeable, though. Which one applies depends on two properties of the work you are about to do next, and to get a better understanding of how to evaluate these properties, we have to take a brief look over two characteristics of an agent: what it sends on each turn, and what a compaction does to it.
We will use Claude Code CLI as the agentic tool of choice here, but pretty much every agentic tool out there faces the same issues and provides similar tools to tackle them. The underlying mechanism is also very similar, so everything you read below should be transferable and worth knowing, independently of which tool you use.
How 2M+ Professionals Stay Ahead on AI
What’s the secret to staying ahead of the curve in the world of AI? Information.
Luckily, you can join 2,000,000+ early adopters reading The Rundown AI — the free newsletter that makes you smarter on AI with just a 5-minute read per day.
What an agent sends on every turn
A model remembers nothing between requests. Everything it knows about your session arrives in the request itself, so the agent re-sends the whole conversation every turn: the system prompt, your project context, every prior message, every tool result, and whatever you just typed. New content goes on the end, so most of each request repeats the one before it.