How to protect your agent's context

Añadido: 2026-05-30

[00:00:00]When you're coding with agents, you're constantly battling context. You give it too much context and it gets confused and sloppy. You give it too little context and it has no idea what the hell you're talking about. So, you got to give your agent a place to write stuff down and hopefully a place where agents can communicate.

[00:00:18]I think I may have solved it. Okay, here we have uh Codeex running inside of Solo. Solo is a meta harness. It's not a coding harness like Claude, Codeex, Open Code, Gemini, AMP, whatever. It sits above those. Okay. So, I'm going to say, can you see the solo MCP? I have added no skills. All I've done is I've enabled the solo MCP. So, this is a codeex off the street. It's nothing special. No system prompt, nothing. So, it says, I see SoloMCP. I called who am I? It says, yes, I can see the SoloMCP. I'm bound to process named codeex running in project 2 which is named desktop one. So this codeex already knows who it is. Okay, what types of things can we do with the solo MCP? I'm going to show you first.

[00:01:09]I'm just trying to gather up some context here. I'm going to show you where we can write that context down.

[00:01:15]Now, normally you're talking to an agent using um some sort of like research skill or interview style skill or maybe you've done deep research on the web and you end up with usually like a huge markdown document, right? Whether it's written in text on the screen or it's an actual downloaded thing from maybe Claude or Gemini Deep Research, you end up with a bunch of context here. I'm just generating context that we're going to save. Okay. So we can use the Solo MCP to manage and inspect Solo itself.

[00:01:46]We can control processes, projects, terminal interactions. We can read the output. We can send control bytes. We can spawn new solo terminals. We can do timers. It's all very cool if I if I may say so myself. Now imagine you're working on a plan. You've come up with a great plan and the context is about to be compacted. Okay, so uh the cliff is coming and you want to write this down. Okay, so use the solo MCP to create a scratch pad with all this info.

[00:02:25]That's all you have to do. You just tell it write this stuff down. And I'll show you uh why having this internal is so useful. Scratchpad API is available. I'm writing it with tags. That's a start.

[00:02:38]It's nice to have tags on things so it's easy to find later.

[00:02:44]And there we go. Created the solo scratchpad. Oh, nice. We got a deep link. That's cool. Let me click over here. Hey, look at this. We've got some nice markdown here. All right. What about what's in desktop? Oh, man. You see, I use scratch pads all the time because I usually do a bunch of research. I I come up with a plan. I form it. I mold it. I make it what I want. And then I say, "Write it down."

[00:03:08]Write it down so that we don't forget.

[00:03:10]So then you can do cool stuff like I'm just going to copy this link and I'm going to open a Claude and I'm going to say, "Hey Claude, can you read this?"

[00:03:19]It'll figure that out. Can you read this guy for me? And I uh No space. That's fine. It'll figure it out. So I said, "Can you read this guy for me?" And I did. I uh pasted the deep link to the Solo Scratchpad. And so look, it already figured it out. This is This is Claude off the street. I didn't do any skills or any anything like that. It says, "Great. Yep, I read it. Here we go."

[00:03:43]Another nice thing that you can do is you can tell your agent, great. Take that scratch pad and turn it into to-dos. Turn it into a bunch of to-dos.

[00:03:53]And then each individual agent gets a little bit of context and it can reference back to the scratch pad. So, we can go over here and we can go desktop one to-dos. and you'll see, oh my goodness, we have so many things we have to do. But each agent can grab one of these to-dos, read the context, and in some cases, it can go, look, this one, this to-do comes from scratchpad 181. Here's the goal. And so you hand this to-do to an agent and you say, "Here is what you're supposed to do. A fresh agent with beautiful context. You hand it this and say, "This is what you're supposed to do. Read the to-do.

[00:04:31]If you want, go read the scratch pad.

[00:04:33]When you're done, update it, close the to-do, leave a comment. This is how you keep your agents operating in tip-top shape, right? You do a huge amount of research. You come up with the perfect plan, and then you break it down into small, discrete slices that you can hand off to fresh agents with fresh context windows, but they can still read all of the reasoning behind it through the to-dos or through the deeper scratch pads. If you like this, check out Solo at solotterm.com. This is a meta harness. You don't have to give up Claude or Codeex or Gemini or AMP or Open Code or whatever. This just adds extra primitives on top that your agents can use to operate better.