QodoAI/Webinar/25/03/Building Qodo with LangGraph
Webinar: From Code Completion to Multi-Agent Coding Workflows - building QodoAI with langgraph
- My Notes
- Says that they would have built Qodo in langgraph if it had been available
- System 2
- Trad vs agentic AI
- agentic - multi-step, autonomous
- they launched qodo gen 1.0 in JetBrains etc
- mentions AlphaCodium
- My Notes
- you made some comment about how langgraph came a day after AlphaCodium
- Flow Engineering stuck with him at that time, he uses that term all the
- My Notes
- original code was AlphaCode by DeepMind
Demo 1 - something simple
- My Notes
Person/Yoav Orlev
- My Notes
- Qodo Gen extension on left in VSCode/Q
- Question he asks it: best web frameworks in python ...
- it generates an "artifact"
- they are printing it as a new file to see it
- you generate the diff after and then apply it to the file
- [to yoav] you should hint at stability in a way that leads to langgraph above
Person/Yoav Orlev
- we like langgraph to leverage the IDE
- this is a structured flow, not a single node
- we're widening this to see if we wanted to have test generated code
- this is first time we're showing this
- are all the tools #MCP based?
- the one we use internal ... follow the same interface
- asks if they actually have same interface or same MCP/Server under hood
- Person/Yoav Orlev
- smaller toolsets are more accurate
- Person/Itamar Friedman
- you can be a bit specific as you learn to prompt
- being a bit verbose with what you want you can get up to 33 tools
- not sure if up to 100 will work well
Demo 2 - #MCP with Playwright
- Person/Yoav Orlev is using ArcBrowser
- doesn't go very well; the chromium browser doesn't pop up
- Person/Itamar Friedman mentions that it usually opens in the background
Discussion of frameworks
- a bunch of discussion here
Challenges with #LangGraph
Things that work well
- LangSmith "I don't think we would manage to get where we are without
Where is it going next
- MCP is like agent to tools, but [something] is like agent to agent
Talking about MCP
- Person/Itamar Friedman
- original documentation about mcp was about fetching context
- there are some problems with using MCP with agent to agent
My Questions from the Discussion
- How does MCP fit into the architecture? Specifically:
- How many servers were created vs. reused?
- How are prompts versioned within MCP server definitions?
- What best practices were learned for versioning MCP servers?
- Is there a good pattern for abstracting custom streaming away from the graph definition in LangGraph?
- LangGraph's integration with the application appears tightly coupled to the graph architecture, limiting flexibility.
- Are there strategies to mitigate this coupling?
- How does LangSmith play into LLM system evaluation with LangGraph?
- How are datasets structured in this process?
- How does the "agent-to-agent" paradigm relate to MCP clients acting as both servers and clients?
- How does this enable chaining agents together while keeping inference controlled by the client application?
- Does Qodo have something similar to "Cursor Project Rules" that allows steering its functionality?
- What are the abstractions around this, if any?
- What best practices exist for managing scoped permissions with MCP servers?
- Other than using finely grained permissions, are there strategies for handling permissions across multiple projects?
- Each MCP server needs its own permissions—how can this be streamlined?
- Is "context credential ops" on Qodo’s product development radar?