Flow Engineering with LangChain/LangGraph and CodiumAI

"Flow Engineering" is a term that has been gaining in popularity recently. The first time it was mentioned as term was in CodiumAI paper on AlphaCodium, where they used flow engineering to produce state-of-the-art results on coding problems.
Flow Engineering can be used for many problems involving reasoning, and can outperform naive prompt engineering. Instead of using a single prompt to solve problems, Flow Engineering uses an interative process that repeatedly runs and refines the generated result. Better results can be obtained moving from a prompt:answer paradigm to a "flow" paradigm, where the answer is constructed iteratively.
"Flow Engineering" is also an incredibly apt name to describe the types of problems that the LangGraph framework seeks to solve. LangGraph is a low level, highly controllable framework for creating agentic and multi-agent applications.
Hear from Itamar Friedman (CEO of CodiumAI) and Harrison Chase (CEO of LangChain) on best practices, insights, examples, and hot takes on flow engineering!

[@BruceChouTW]

This is the best agentic workflow interview i’ve ever watched. thank for sharing.

[@Bakobiibizo]

I've reached a similar conclusion about LLMs and current approaches to writing agents. I've collaborated with two research groups to develop multimodal agents, but we encountered an issue of decoherence when loading too many decision-making processes and providing the agent with excessive options. I've found the most success with extremely narrowly scoped, finely-tuned models designed for specific steps in the required sequence. I use a state machine to manage the transition to the next step; the agent indicates when it believes it has completed its task, and then an evaluation model confirms the task's completion. Whenever possible, it's beneficial to offload extraneous context and decision-making to programmatic systems. I could envision an orchestration model that makes high-level decisions, but that seems questionable to me still.

[@DougKingPDX]

Orchestration works well. Putting LLM classifiers in the orchestration steps to understand context, intent works well. Discrete steps in workflows allow for better guardrails and direction conforming to the workflow requirements.

[@XShollaj]

Check Dify for quick prototyping

[@XShollaj]

This is great. Figuring out efficient agent architectures is the next big step.

[@mr.daniish]

Great guest and tons of value in the talk!

[@DonBranson1]

Great video. I hope Itamar is brought back for a follow-up discussion on LangGraph. I've found that getting the LLM to generate and iterate on a mermaid.js model can lead to better flow engineering.

[@jonclement]

Good point. yaml as a nested dict misses links, markdown for txt layouts, but to fuel LLMs need for text descriptions of ER diagrams etc -- mermaid seems like a good choice.

[@kenchang3456]

I agree, this is a great interview and am definitely interested in its premise. Thank you for sharing.

[@IkechiGriffith]

Amazing stuff. Thank you

[@_ptoni_]

great talk, ty

[@AGI-Bingo]

Fantastic Interview! Hope we get a chance to have you both on the channel ❤ doing an ai podcast segment for agentic flow developers.
#WholesomeAGI coming soon

[@chandrurulz82]

Good job sir

[@Aripb88]

Thank you for such an in depth discussion! So is the next step having models train on repos of flows? Ultimately you are left with an input/output problem with a performance metric 😂

[@elimelethlopez6859]

luv Langchain ❤

[@pnhbs392]

What is your definition of "Flow Engineering?"