The recent conversation between Andrew Ng and Harrison Chase at the LangChain conference provided deep insights into the current state and future direction of AI agent development. Let's dive into the key takeaways from their discussion and explore the essential points that agent developers need to know.
A Fresh Perspective on "Agentness"
Andrew Ng's concept of "agentness," introduced a year and a half ago, remains highly relevant. Instead of asking "Is this an agent or not?", he suggests approaching it as a spectrum of how autonomous a system is. This perspective has become even more important now that marketers have started throwing around the term "agent" liberally.
According to Ng, most current business opportunities lie not in highly autonomous systems with complex loops, but in workflows that are linear or have slight branching. Think tasks like checking website forms, performing web searches, and verifying compliance issues in databases.
Core Skills Every Agent Developer Must Master
The Critical Importance of Evaluation Systems
Andrew Ng points out that many teams start building systematic evaluation systems far too late. Most teams rely on manual result checking for way too long whenever they make changes to their systems.
An effective evaluation system not only helps you understand your overall system performance but also tracks individual steps so you can pinpoint exactly what's going wrong. He advises thinking of evaluation not as a massive project, but as a simple tool you can whip up in about 20 minutes.
Leveraging Tools Like Lego Blocks
Ng compares AI tools to Lego blocks of different colors and shapes – the more types of tools you know, the faster you can assemble something amazing. Whether it's RAG, chatbot building, memory systems, or guardrails, having a diverse toolkit means you can quickly pick the right tool for the job.
However, these tools and best practices are constantly evolving. For example, as LLM context lengths have grown, RAG best practices from a year and a half ago have changed significantly. Where we once used complex recursive summarization techniques, we can now stuff much more information directly into the context.
Underrated Technologies Worth Watching
The Untapped Potential of Voice Stack
Andrew Ng believes voice applications are massively undervalued. While large enterprises are showing tremendous interest in voice applications, the developer community's attention is relatively limited.
The biggest advantage of voice is reducing user friction. Text prompts can feel intimidating to many users, but with voice, you can naturally continue the conversation as time flows. People feel less pressure to be perfect when speaking compared to writing, making them more willing to share information.
The biggest technical challenge for voice applications is latency. You need to respond within one second of the user speaking, ideally within 100 milliseconds. This is where "pre-response" techniques come in handy – using phrases like "That's interesting" or "Let me think about that" to mask the delay.
The Democratization of AI Coding Assistants
The productivity gap between developers who use AI assistants and those who don't is enormous. Yet many company CIOs and CTOs still maintain policies prohibiting engineers from using AI assistants.
Interestingly, everyone at AI Fund – including the receptionist, CFO, and legal team – knows how to code. Not to become software engineers, but to give more precise instructions to computers in their respective roles.
Emerging Standards and Protocols
The Rise of MCP (Model Context Protocol)
MCP represents an important attempt to standardize integration between agents and various data sources. The goal is to make it possible to integrate n models with m data sources with n+m effort instead of n×m effort.
Currently, MCP is in its early stages. Many MCP services you can find online don't work properly, and authentication systems are unstable. Also, the current approach provides long lists of available results, but we'll need hierarchical discovery mechanisms in the future.
Inter-Agent Communication Is Still Premature
While there's high interest in multi-agent systems, we're actually still in very early stages. It's possible for one team to build multiple agents that communicate with each other, but there aren't many successful cases of agents built by different teams communicating effectively.
Vibe Coding and the Future of Development
The term "vibe coding" is misleading. It sounds like coding based purely on intuition, but it's actually a highly intellectual activity. After spending a full day coding with an AI coding assistant, you'll be genuinely exhausted from the intense concentration required.
Advising people not to learn coding because AI will automate it might be one of the worst career advice in history. Historically, whenever coding became easier, more people started coding. We had similar concerns when we moved from punch cards to keyboards, from assembly language to COBOL, but ultimately we needed more programmers.
One of the most important future skills will be the ability to give precise instructions to computers. Understanding how computers work allows you to create more accurate prompts and instructions.
Key Factors for Startup Success
Based on AI Fund's experience, the factors for predicting startup success are:
First is speed.The execution speed of experienced teams is unimaginable to those who haven't witnessed it.
Second is technical knowledge.While knowledge of marketing, sales, and pricing is important, this knowledge is relatively widespread. Deep understanding of how technology actually works is rare and valuable.
Wrapping Up
Agent development is still in its early stages, but there are many opportunities to create practical business value. Rather than building complex autonomous systems, it's more realistic to start by automating the linear workflows that people are currently performing.
For successful agent development, the keys are building systematic evaluation systems, understanding various AI tools, and executing quickly. You should also pay attention to emerging technologies like voice interfaces and MCP.
Most importantly, instead of fearing that AI will replace coding, focus on developing the ability to work more effectively with AI. The ability to give precise instructions to computers will become increasingly important, and basic programming knowledge will still be necessary for this.
The future belongs to those who can bridge the gap between human intent and machine execution. Whether you're building simple workflow automation or exploring the frontiers of autonomous systems, the fundamental skills of evaluation, tool mastery, and clear communication with machines will serve you well in this exciting field.