Erick Friis

Engineer and expert on agentic AI engineering at LangChain

Mentions

Appeared in

Discussed in

Key positions and views

The single biggest hurdle to productionizing LLM applications is reliability, due to the non-deterministic nature of model outputs, which necessitates extensive guardrailing.

Agentic architectures built as state machines with LangGraph are superior to simple agent loops because they offer more control, reliability, and the ability to handle complex, cyclical workflows with human oversight.

A 'less is more' development philosophy is most effective for building AI agents, favoring simple key-value memory, fewer opaque abstractions, and limiting an LLM's access to about five tools at any given step.

Tool calling is the most critical LLM feature for practical application development, as it provides the essential structured interface between probabilistic models and deterministic code.

A hybrid model strategy, using smaller models for simple tasks like classification and larger models for generation and control flow, is a key optimization for managing cost and performance.

Podcast consensus on Friis

Points of consensus

▶Erick Friis consistently identifies reliability, stemming from the non-deterministic nature of LLMs, as the primary obstacle to production deployment.May 2026

▶He advocates for structured, stateful agent architectures like LangGraph over simpler, less reliable agent loops, especially as complexity and the number of tools increase.May 2026

▶He repeatedly emphasizes a 'less is more' approach to agent design, recommending limiting the number of tools available to an LLM at any given time and using simple memory structures.May 2026

▶He views tool calling as the single most important LLM feature for developers, as it creates the essential structured bridge between probabilistic models and deterministic code.May 2026

Points of debate

▶Friis highlights a key tension between the excitement for powerful, multi-modal models and their current lack of reliability for production use.May 2026

▶He describes a conflict in agent design between giving an LLM many tools for flexibility and the performance degradation that occurs when too many tools are available.May 2026

▶He points out an inconsistency in the market where the performance of open-source models like Llama can vary significantly depending on the hosting provider's implementation of features like tool calling.May 2026

▶There is an ongoing strategic choice for developers between using large, expensive models for all tasks versus optimizing costs with a mix of smaller, specialized models for simpler tasks.May 2026

Key themes

▶Pragmatism in AI EngineeringMay 2026

Friis consistently emphasizes the practical challenges of productionizing LLMs, focusing on reliability, non-determinism, and the need for guardrails over celebrating raw model capabilities. He stresses that current LLMs cannot be trusted 'out of the box' for important decisions.

This suggests the market is maturing from pure R&D to demanding robust, predictable enterprise applications, creating opportunities for tools that enforce structure and control.

▶The Shift to Agentic State MachinesMay 2026

He details LangChain's strategic pivot from simple agent loops, which perform poorly with many tools, to LangGraph. This more structured approach orchestrates agents as state machines, allowing for more complex, cyclical, and controllable workflows, including native 'human in the loop' interruptions.

The evolution from linear chains to cyclical graphs indicates that the complexity of AI applications is growing, requiring more sophisticated orchestration frameworks beyond simple API calls.

▶Economic Optimization of LLM UsageMay 2026

Friis highlights the economic pressures of using LLMs, discussing the significant drop in inference costs and strategies customers use to manage expenses. This includes using smaller models for routine tasks and betting on future cost declines.

The cost of inference is a primary driver of architectural decisions, suggesting that solutions enabling efficient model usage, like LangGraph, have a strong competitive advantage.

▶Developer-Centric Framework DesignMay 2026

He explains LangChain's evolution towards a 'bare bones' philosophy with fewer opaque abstractions in favor of raw Python and JavaScript. This is a direct response to developer feedback for more intuitive and less 'magical' tools that are easier to debug.

The AI developer tooling space is highly competitive, and winning developer adoption requires prioritizing transparency, control, and ease of debugging over black-box abstractions.

Source episodes

Sentiment over time

Not enough data for timeline

Changes over time

October 2022

The LangChain open source project is started by Harrison Chase.

First half of 2023

The company LangChain is officially formed.

Mid-2023 (implied)

Experience with early agent implementations, like the simple ReAct loop, reveals performance degradation as tool complexity increases, signaling the need for a new architectural approach.

Late 2023 / Early 2024 (implied)

LangChain's primary focus shifts to LangGraph, a more robust framework for orchestrating AI agents as state machines to address reliability and control issues.

Present

LangChain's development philosophy has evolved to be more 'bare bones' to improve developer experience, and key features like 'human in the loop' interruptions are now natively supported in LangGraph.

Suggested prompts

How does LangChain's pivot to LangGraph reflect broader trends in the maturation of AI applications from prototypes to production systems? &nearr;Erick Friis identifies reliability as the main hurdle for LLMs. What specific architectural patterns and guardrails, beyond limiting tools, does he suggest for mitigating non-determinism? &nearr;What are the economic implications of the hybrid model strategy (mixing large and small LLMs) that Friis describes, and how might this impact the competitive landscape of model providers? &nearr;Given the inconsistencies in tool-calling performance across providers for models like Llama, what is the role of frameworks like LangChain in abstracting away or normalizing these differences for developers? &nearr;

Key concepts

LangGraph 1 ep LLM Reliability 1 ep Tool Calling 1 ep Agentic Architectures 1 ep State Machines 1 ep LLM Costs & Optimization 1 ep Developer Experience 1 ep Human in the Loop 1 ep Agent Memory 1 ep

Notable quotes

“I would probably say reliability is the first one... In this case, I think it's probably more the outputs are non-deterministic.”

Erick Friis · LangChain and Agentic AI Engineering with Erick Friis

“tool calling is easily the most important feature that Langchain and LangGraph users are using out of the models, where it's really useful for providing some sort of tool calling and structuring output where you provide a schema and the LLM generates all the fields that you kind of ask for, which is a really nice interface point between code and these LLMs.”

Erick Friis · LangChain and Agentic AI Engineering with Erick Friis

“LangChain's development philosophy has evolved to be more 'bare bones', favoring raw Python and JavaScript to reduce opaque abstractions and make the framework more intuitive for developers.”

Erick Friis · LangChain and Agentic AI Engineering with Erick Friis

“A recommended best practice for building reliable AI agents is to limit an LLM's access to approximately five tools at any given step in a workflow.”

Erick Friis · LangChain and Agentic AI Engineering with Erick Friis

Report last updated: May 5, 2026

Get started free

Back to Entities Intelligence Report