LangChain/YouTube/25/08/Introducing OpenSWE
Properties
- date-created
- 2025/08
Video
00:00 Introduction to Open SWE
What is Open SWE?
- Open SWE is an async cloud-based open source AI Coding agent
- Core functionality:
- "you connect your GitHub account to open send it a task and it does the rest"
- "It plans, executes the plan, writes code, runs tests, runs your different scripts"
- "reviews the code before putting up a pull request to make sure it's all high quality code"
Demo Task
- Express.js app enhancement:
- "add an in-memory data.json store to store my data"
- "implement this, write some tests, and then handle all the errors gracefully"
00:50 Cloud-Based Architecture & Benefits
Sandbox Environment
- Powered by Daytona cloud infrastructure
- Key advantages over local agents:
- Unlimited parallel tasks: "run essentially unlimited open suite tasks at the same time"
- Workflow integration: "I wake up, I plan out what I want to get done in the morning. I create all these different GitHub issues for that and then I simply kick off Open Suite tasks for them all at the exact same time"
Security & Isolation
- GitHub integration: "you simply add a label to the GitHub issue and that kicks off a new Open Suite task"
- Sandbox benefits:
- "I don't need to worry about any sort of malicious command, say deleting my entire computer"
- "We don't have any sort of guardrails around the commands it can execute"
- "It doesn't have access to secrets. We don't need to worry about it leaking those"
- "if it wants to execute a shell command that says, you know, delete the entire repository, sure, because it's on its own branch"
02:18 Planning & Execution Workflow
Plan Generation & Approval
- Automated planning: "it's going to start generating the actual plan now"
- User control: "once the plan has been generated, it's going to propose it to me where I can then go and accept it as is"
- Plan modification options:
- "I can edit individual plans, delete them, or even send a message back to the manager"
- "Hey, I actually want you to update the plan in this way"
Real-time Communication
- Manager system: "This is the manager on this side"
- Dynamic updates: "I can send it a message and say actually you know rename this file to XYZ or you know what I've changed my mind on the product spec"
- Advantage over other agents: "with a lot of current coding agents that are that are being used out there, you can't really send it messages and update its task in real time"
03:04 Internal Usage & Success Stories
Open SWE Self-Development
- Self-contribution: "using Open Suite to write Open Suite"
- Repository insights: "it's the second biggest contributor to the open suite repository"
- Use cases: "writing changes to the front end, updating the actual agent documentation"
LangGraph Development
- LangGraph contributions: "tied as the biggest contributor to the Langraph open source repository"
- Complex repository handling: "highly complex repo with many different subpackages, apps, documentation, highly complex tests"
- Team comparison: "tied with Sydney, who's one of the core maintainers of Langraph as the most contributions for this week"
Simple Task Example
- Community issue resolution:
- "Hey, there's a problem in Open Platform. Can you fix it?"
- "I simply added this label to the issue, Open SW Max Auto"
- Result: "very simple oneline change. It figured out exactly what to do and in the simplest possible implementation"
- Quality: "Its very first try worked as is pass CI every single time"
Complex Task Example
- Bench evaluations implementation:
- "extremely complex 19step plan"
- Web research capability: "I gave it links to the actual Swedbench website because open suite can go and research the web for you"
- Results: "eight files, 2600 changes, a nice body description, only commits from Open Suite"
- Success: "All of them pass CI. And at the bottom, we can see everything worked"
06:26 Why We Built Open SWE
Market Gap Analysis
- Missing features: "there wasn't really an open source coding agent that had the same level of features and integrations that open has out there"
- Comparison with existing solutions:
- Local agents: "local open source coding agents like Coder"
- Closed source: "closed source fully autonomous coding agents like Devon or even Cursor"
- Missing integration: "there wasn't anything where I could connect it to my GitHub account add a label to the GitHub issue or invoke the agent through the UI"
LangGraph Specialization
- LangGraph expertise: "we want one that's really good at langraph"
- Current development: "we're working on a version of open suite which is prompted specifically for langraph code"
- Future plans: "open suite langraph engineer. You'll be able to use this to build agentic applications or just agent services for you fully autonomously"
- Agent architecture focus: "even if modern-day coding agents can write langraph code, they typically aren't good at writing actual agents themselves"
08:53 Getting Started & Token Management
Access & Setup
- Open source: "it's an open source repository you simply go to langchainai/openswuite"
- Documentation: "docs.langchain.com/labs"
- Self-hosting: "detailed development setup guide so you can fork open suite use it for your own internal use cases"
Hosted Service
- Easy setup: "simply inputting an anthropic API key"
- Managed infrastructure: "We manage the GitHub app, the agent deployment, the sandbox, everything else for you"
Cost Transparency
- Token tracking: "You get insights into all the different tokens that it consumes"
- Caching benefits: "we're caching all these tokens are anthropic. So only 5.7% of the tokens we've sent to it have been uncached"
- Cost savings: "we've only spent $62 and $222 of that we would have had to pay if we hadn't cached these tokens ahead of time"
- Model breakdown: "We can also look at model breakdown"