Transcription
It's been about a month since our last update video on Klein, where we showcased the 3.18 release. Since then, Klein has leveled up quickly, becoming one of the most capable autonomous AI coding agents out there. For those who are unfamiliar, Klein is an open-source autonomous coding agent that lives directly within your IDE, capable of creating, editing files, executing commands, using the browser, and so much more.
Now, with the release of version 3.25, Klein introduces many new powerful features that make it even better at tackling long complex tasks completely on its own. From new agents to focus modes and much more. This update is packed. So, let's just simply dive in. Now, just take it in. The last update video I made was a month ago, and you can see so many new updates that have been introduced to Klein since that update. There has been approximately a couple of pages worth of new edits that have been implemented to Klein, making it even better.
First up is the terminal fix. This is where terminal bugginess has been a long-standing issue that Klein used to face with its generations, which would prevent Klein from actually getting full context of it whenever it's executing any sort of commands. But with this new fix, it's simple but brilliant because within the shell, it has it so that it integrated the API fails to capture output, and Klein can now automatically fall back to reading the terminal contents directly using the same mechanism behind the terminal output feature. And this is where you don't have any sort of blind spot. Klein can now see exactly what you see in the terminal, no matter the shell integration status.
Next up is where Enthropic recently had made a huge release where Claude Sonnet 4 now has a 1 million token context window, and Klein can now fully actually support it. That's five times more than its previous 200k limit. And it fundamentally changes how you can actually work in Klein, which can unlock deeper planning, richer context, and full end-to-end code-based reasoning without compromise. Because of that, it has improved Klein in two ways. You first have it so that there is better code quality, where you don't need to actually be stingy with the context. It is going to be capable of pulling in every file, document, and MCP provided context you need. With Sonnet 4's 1 million token window, Klein can now see the entire picture of your codebase, making the coding decisions a bit better than previously. The second way it improves Klein is with longer and more complex sessions. Those iterative cycles of code, test, and refining can now be much better because of longer context without losing the actual context as you're generating with Klein. No more restarting tasks mid-feature because you hit a token limit. This is where this new larger context is going to be able to help you generate better with the long planning sessions. Now, even though there's prompt caching, just keep in mind it gets more expensive over the 200k tokens. So, this is where you want to keep this in mind when you're generating with Klein, because if it's autonomous and it keeps on making edits over and over again, it could tally up a huge API cost.
AI is rewriting the rules of business, and BoxWorks 2025 is where the future comes alive. Join top leaders from OpenAI, Google, Service Now, and Box. Explore how AI and agents can unlock your unstructured data, extract insights, and automate workflows. All with next-level security. Learn directly from the innovators shaping the AI-first era of business. Whether you're a CIO, developer, or just starting your AI journey, BoxWorks delivers keynotes, deep-dive master classes like "Discover the Latest in Content Plus AI," diving into tailored sessions and tracks, and so much more. You're going to be able to get hands-on experience where you can put it into action immediately. Connect with peers, get inspired by industry icons, and see game-changing AI in action. September 11th to 12th at the San Francisco Marriott Marquis, in person or virtually. This is your chance to thrive in the AI-first era. So, make sure you register now with the link in the description below and use my code for 50% off. Don't just watch the AI revolution, lead it with Box at BoxWorks 2025.
But specifically talking about the version 3.25, it is tackling one of the biggest challenges in AI coding, where it keeps agents on track through long and complex tasks. This update introduces a three-part system, three interlocking features that work together to keep Klein focused, consistent, and effective end-to-end. With Klein, this thread is never lost. And the main focus as to why they developed this is because they wanted to tackle the longest and most complex coding tasks that often trip up other agents. This is where they found that large language models usually struggle as context grows. And this is obviously due to "lost in the middle," which is a phenomenon that usually happens with AI agents, as well as multi-turn interactions that make things worse, causing accuracy to degrade over time. And simply expanding context windows wasn't actually enough, since signal-to-noise ratios were actually deteriorating the actual response.
So, to address all of this, it introduced a three-step interlocking system that keeps agents focused and on track, allowing you to fully leverage frontier models and maintain performance even on ultra-long multi-step projects. This is where it first starts off with the deep planning phase, where you run the deep planning trigger, and it sets up a four-step workflow where Klein silently analyzes your codebase, as targeted classifying questions, as well as generating a comprehensive implementation plan as a markdown within your root directory. It then hands off the next task with all the relevant tasks as well as a plan included. And overall, this is going to ensure your next agent starts with the perfect context and a clear path forward.
Then it moves on to the second phase, which is the focus chain, where your persistent north star is going to be guiding your agents. This is where it's going to keep Klein on track by generating a to-do list from your task and reinjecting it into the context at regular intervals, every six messages by default. It updates the checklist as you work and it's going to check off completed items, adding new tasks, as well as shifting priorities, and overall, it's going to ensure the core mission never gets lost despite context noise or multi-turn degradation.
And then lastly is the autocompact, compress, and continue phase. This is when context limits are actually reached. The autocompact will automatically summarize the entire conversation, technical decisions, code changes, as well as the progress, and replace the bloated history with the compressed summary. The task then continues seamlessly. And with an autocompact feature, even tasks that require millions of tokens can be completed in a fraction of the context window. Combined with deep planning as well as the focus chain, this is where Klein can become a coding agent that's capable of tackling the most complex long-running projects with precision and efficiency, which is where you're going to save a lot of time as well as save a lot of resources in terms of your token expenditure.
There's a couple of other small things a part of this update, like the added 200k context window support for the Claude Sonnet 4 in OpenRouter and Klein providers, added custom basel for requesty, uh, the fixed the duplicate attempt completion command in progress, there is fixed bugs preventing announcement banner to be dismissed, and added the GPT OSS model on to the AWS Bedrock provider. Now, there's also a lot of other small updates and fixes that are implemented within the overall GitHub release page. So, if you want, you can take a look at this, and you can get a good comprehensive understanding of all individual small updates.
And if you go into the settings, you'll see a couple of other smaller things that will showcase smaller updates, like configuring the terminal browser settings, as well as smaller features that have been implemented for each individual model. This is a small comprehensive UI fix that has been implemented within Klein itself, making it more usable and user-friendly.
If you like this video and would love to support the channel, you can consider donating to my channel through the Super Thanks option below. Or you can consider joining our private Discord, where you can access multiple subscriptions to different AI tools for free on a monthly basis, plus daily AI news and exclusive content, plus a lot more.
But that's basically it, guys, for today's video on the new Klein update. I'll leave all these links in the description below so that you can easily get started. But with that thought, guys, thank you guys so much for watching. I hope you enjoyed today's video and got some sort of value out of it. Make sure you go ahead and subscribe to the second channel, join the newsletter, join the Discord, follow me on Twitter, and lastly, subscribe, turn on the notification bell, like this video, and please take a look at our previous videos, because there's a lot of content that you will truly benefit from. But with that thought, guys, have an amazing day, spread positivity, and I'll see you guys fairly shortly. He suffers.