AI negotiators close $8 billion in deals for Fortune 500

Plus: a course on memory for agentic systems, a new vision language model for GUI agents, and more

Author

Michael Cunningham
November 11, 2024

Welcome back to Building AI Agents, your biweekly guide to everything new in the AI agent field!

Read Building AI Agents to improve your professional life—and your love life

In today’s issue…

  • Walmart uses AI agents to negotiate $ billions in deals

  • Microsoft’s overhauled AutoGen framework is out

  • A course on adding memory to your agents

  • A foundation model for visual agents

…and more

🔍 SPOTLIGHT

Created using DALL·E 3

If your negotiating partner in your next business deal starts their opening offer with “As a large language model, I cannot…”, you might be dealing with an AI negotiating agent.

Major corporations such as Walmart and the shipping industry heavyweight Maersk are now beginning to power their contract negotiations with AI agents built by Estonian startup Pactum, which recently raised $20 million in venture funding. The company’s technology, which it refers to as an “autonomous negotiating suite”, is based on a custom large language model trained on business contracts and market data, giving it the necessary domain knowledge to understand the subtleties of dealmaking. When provided with the necessary context, it can draw up contracts for a vendor or autonomously negotiate with the vendor itself, including responding to counteroffers. Walmart and Maersk now use Pactum’s agent to negotiate for a combined total of over $8 billion in spending, achieving tens of millions of dollars in savings.

Pactum is not the only company offering agentic AI negotiators—in November of last year, UK-based Luminance demonstrated the ability of its Autopilot program to hammer out the details of a legal contract. While the company already offers a simpler chatbot known as Lumi designed to assist lawyers in negotiations but not to directly participate itself, Autopilot is intended to fully automate the process. “This is just AI negotiating with AI, right from opening a contract in Word all the way through to negotiating terms and then sending it to DocuSign,” said the company’s chief of staff and managing director. Luminance, too, already counts major enterprises such as Staples and Koch Industries among its clients.

Contract negotiations consume significant amounts of time from lawyers and other highly-paid professionals, making them a major financial pain point for companies which must maintain relationships with large numbers of suppliers or clients. As they require significant soft skills and the ability to reason about the effect of many different pieces of information in context, they were previously outside the abilities of artificial intelligence. However, with the arrival of large language models, companies are now gaining the ability to outsource this work to AI systems which can work tirelessly 24/7 at minimal cost.

A new survey of procurement heads by tech consulting firm Gartner found that 58% are already implementing or plan to implement AI solutions. By 2027, the firm predicts, 50% of organizations will engage in AI-powered contract negotiations. As with many other fields, the future of business negotiations is looking increasingly agentic.

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

Refer one person and we’ll send you a set of weekly bonus links to additional AI agent news for a month! See a sample here

🤝 WITH AI TOOL REPORT

There’s a reason 400,000 professionals read this daily.

Join The AI Report, trusted by 400,000+ professionals at Google, Microsoft, and OpenAI. Get daily insights, tools, and strategies to master practical AI skills that drive results.

📰 NEWS

AutoGen conversation pattern | Source: AutoGen

Microsoft’s AutoGen multi-agent framework released an early version of its highly-anticipated new architecture after a complete overhaul.

Agent startup Fetch.ai’s Fetchai Innovation Lab collaborated with the UN Convention to Combat Desertification and the Chan Zuckerberg initiative to host a hackathon for users to build agentic solutions to global challenges.

The CRM giant is hiring 1,000 workers to sell its AI agent product Agentforce, which the company says has drawn significant interest and positive feedback since its launch two weeks ago.

🛠️ USEFUL STUFF

Letta founders Charles Packer (left) and Sarah Wooders (right) | Source: TechCrunch

A new DeepLearning.AI course in which the founders of agent memory startup Letta show how to use the company’s tech to build an LLM agent with long-term memory.

Composio AI has released SWE-Kit, an open-source toolkit and headless IDE allowing users to build custom coding agents. The company’s agent built using the kit achieved SOTA performance on the SWE-Bench AI software engineering benchmark.

Replit is hosting a tech talk and social at their San Francisco office on how their much-hyped AI software engineer was built, and what it taught them about the art of creating multi-agent systems.

A tutorial video by Dendrite Systems showing how to build an agent that can use a headless browser to interface with websites, including those difficult to access for traditional web scrapers.

An X thread on how a Lindy.ai user automated his appointment scheduling, email replies, note-taking, and more with the company’s agentic assistants.

💡 ANALYSIS

Source: Freerange Stock

An interview with automation expert Phil Fersht, in which he discusses the potential for corporate chief innovation officers (CIOs) who become early adopters of AI agents to reap outsized rewards, and lays out the technical challenges they will need to overcome.

This piece examines how AI agents are automating the cryptocurrency world, shows how to build a crypto agent with OpenAI’s Swarm and Coinbase’s developer platform—and warns about the perils of the recent agent memecoin trend.

🧪 RESEARCH

Operating modes of OS-Atlas | Source: arXiv

OS-Atlas is a new vision-language model (VLM) trained from the ground up on over 13 million GUI elements to allow it to power agents capable of interacting with graphical interfaces.

SCIPE is a tool built by Berkeley researchers for identifying the failure node in a chain of LLM calls, allowing users to diagnose and fix problems in complex agentic systems.

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

What did you think of today's issue?

Login or Subscribe to participate in polls.

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

Follow us on X (Twitter), LinkedIn, and Instagram