Replit's software engineering agent stuns users

The company's new AI programmer shows impressive capabilities

Author

Michael Cunningham
September 09, 2024

Welcome back to Building AI Agents, your weekly guide to everything new in the AI agent field! We hope you like the new format!

Happy Monday to all you agent builders, except for whoever is apparently using them to start internet cults

In today’s issue…

  • Replit’s new AI agent builds apps in minutes

  • Complex societies emerge from massive multi-agent simulations

  • A roundup of recent agent papers

  • A podcast on the new era of agents with the founder of Sierra AI

  • A MOOC on agent building

…and more

🔍 SPOTLIGHT

Source: Replit

Once again, an AI software engineer is the talk of the town.

From Cognition’s Devin—and its open-source competitor OpenDevin (now OpenHands)—to Devika and more, a variety of companies and open-source projects are seeking to go beyond simple AI-powered code completion to create software engineering agents capable of performing many of the same tasks human SWEs can, such as creating unit tests and documentation, fixing bugs, and even building entirely new pieces of software from scratch.

The latest entrant in the game is Replit, a San Francisco-based startup which provides an online IDE for collaborating on software projects. On Thursday, the company lit the tech internet on fire with its release of Replit Agent, an AI software engineer which—according to many excited users breathlessly posting their demos—allowed coders with little or no experience to build entire functional apps in minutes using only natural language prompts. Even legendary OpenAI and Tesla researcher Andrej Karpathy was impressed, declaring that that Replit’s tool earned a spot in the “feel the AGI” category, and pointing out that its ability to step beyond merely writing code to automate the entire engineering process would allow even novice users to develop and deploy whole applications.

Replit Agent is currently in early access and available to Replit users subscribed to Core or Teams at no additional cost, though the company says that further pricing information will come later in the year. As users continue to push the boundaries of its capabilities, they will undoubtedly discover that the software has its limitations, as all AI agents do. Software engineers will not see their jobs replaced quite yet, but Replit Agent—as with Devin before it—provides another glimpse of an impending future in which they will increasingly serve as higher-level managers of the actual hands-on coders: AI agents.

If you find Building AI Agents valuable, forward this email to a friend or colleague!

📰 NEWS

Source: Reddit

Using the video game Minecraft as a model, the startup Altera found that a simulated community of over a thousand AI agents resulted in the emergence of democracy, religion, and a commercial economy. While these simulations may seem frivolous, video games like Minecraft have become a common arena for training agents to function in human-like roles.

Coval, a new startup backed by Y Combinator, has launched with the goal of providing agent builders with the ability to automate simulation and evaluation of their products. Its founder, Brooke Hopkins, formerly headed a similar testing unit at self-driving car company Waymo.

LangGraph, one of the primary agent frameworks, is now available in JavaScript, presenting an option for developers outside of the traditionally Python-dominated agent ecosystem.

Agent startup RevDev closed a sizeable $100 million funding round, valuing it at $1.1 billion—officially a “unicorn” in Silicon Valley lingo. RevDev provides a platform called AgentOS, which allows enterprise customers to build agents on top of a knowledge graph of their organization’s data.

API costs for OpenAI’s leading models have fallen to 1/240th of those at the release of GPT-4 in March 2023, a more than 100-fold greater rate of change than Moore’s law.

🧪 RESEARCH

Source: arXiv

AI writer and influencer Rohan Paul presents a list of the most influential agent papers.

This paper provides the technical report for Salesforce’s recently-released xLAM large action models. Unlike traditional LLMs, large action models are purpose-built for agentic workflows.

As this week’s Replit release demonstrates, software engineering agents are an extremely hot area. The authors of this paper present a review of over 100 papers in this domain, including the different approaches they take and the challenges and future of the field.

AI agents traditionally utilize frontier-class LLMs running in centralized datacenters, but this paradigm is not applicable to all use-cases. This paper introduces a framework for training and deploying small language models running on edge compute to power agentic applications.

🛠️ USEFUL STUFF

Source: Wikipedia

Bret Taylor has played key roles in some of the most significant tech companies, including co-creator of Google Maps, CTO of Facebook, co-CEO of Salesforce, and chairman of the boards of Twitter and OpenAI. Now, as the founder of conversational AI agent startup Sierra, he speaks about his career and the enormous potential of agents.

Dawn Song and Xinyun Chen of UC Berkeley and Google DeepMind, respectively, present a new MOOC on LLM agents, covering infrastructure, design, evaluation, ethics, and more—replete with guest speakers from many of the most important institutions in the field.

Microsoft hosts an internal conference on AI agents 4x per year, according to employee, influencer, and “Chief Troublemaker” Dona Sarkar. Though this event is open only to Microsoft employees, Sarkar says that an external one is coming soon.

💡 ANALYSIS

Source: SmartBrief

Hyper-specialists currently command the most demand—and highest salaries—in the tech world, due to their deep expertise in narrow technical domains. This article makes the case that the rise of agents will shift power to hyper-generalists who can oversee teams of highly specialized automated agents.

The author of this article contends that the first generation of agents, built around the ReAct framework, were overly broad in their goals and often failed. The second generation, she says, will be more narrowly tailored to specific goals, and will fall into one of a set of distinct architectures.

With the xLAM action models described and its Agentforce platform, Salesforce has made a “hard pivot” to AI agents, according to its founder and CEO Marc Benioff. This piece analyzes the impact Salesforce’s agents—and generative AI more broadly—is having on sales organizations.

Thanks for reading! Until next time, keep learning and building!

If you have any specific feedback, just reply to this email—we’d love to hear from you

Follow us on X (Twitter) and LinkedIn