- Building AI Agents
- Posts
- Microsoft announces Copilot Studio agents
Microsoft announces Copilot Studio agents
Developers will soon be able to build customized LLM agents for business tasks through the company's Copilot Studio
🔍 Spotlight
At Microsoft’s annual Build conference on Tuesday, the company announced the addition of new agentic capabilities to its Copilot Studio platform. Previously, Copilot’s chatbots could only respond to direct user queries, but they will soon be able to activate in response to events such as emails while remembering prior interactions, allowing sophisticated workflows to be run without human input.
Copilot was launched in February 2023, enabling business users and their customers to converse with chatbots powered by OpenAI models, which, in turn, could take simple actions to help answer the users’ queries. However, these systems were intended largely to be simple question answerers, rather than fully-fledged autonomous agents. The new copilots promised at Build, however, will be able to initiate long chains of actions and remember details of prior events, enabling them to automate workflows that occur asynchronously over the course of days. Microsoft offered a variety of example use cases, including managing IT help tickets from beginning to end by contacting the relevant support personnel, obtaining manager approvals for fixes, and closing the ticket when resolved.
Copilot’s new agentic features are currently only available to select users participating in a private preview, but Microsoft plans to roll them out more broadly later this year. As with OpenAI and Google’s competing unveiling of agent systems earlier this month, Microsoft’s announcement reflects the rapidly growing interest in autonomous agents among the largest players in AI.
đź“° News
From May 15 through June 17, Nvidia and LangChain are hosting a contest to build agent tools powered by the two companies’ LLM hosting and integration products. The contest’s two winners and special mention runner-up will receive Nvidia GPUs, with various other prizes available to the remaining participants.
Adept, a company launched by Google alums Ashish Vaswani and Niki Parmar to build visual agents capable of performing sophisticated workflows on users’ desktops, has spoken with major tech companies about a sale. The startup was recently valued at at least $1 billion, but has been rocked by the departures of Vaswani and Parmar and rumors of infighting.
Cognition Labs’ software development agent Devin will soon be available on Microsoft’s Azure cloud, the company announced at its Build conference. Devin’s announcement last month created both excitement and controversy with its promise of automating software engineering jobs, though many have claimed that its capabilities are exaggerated.
The boldly-named Amazon AGI unit is forming a new team focused on LLM agents, according to a LinkedIn post by Amazon director Shiv Vitaladevuni, the team’s head. Vitaladevuni is seeking applicants with experience building agents for a variety of technical roles.
H, formerly known as Holistic, has raised a massive $220 million from prominent tech billionaires and venture capitalists to advance its AI agent capabilities. The company promises to build “full AGI” that can automate business processes.
🧪 Research
Google DeepMind researchers present ANDROIDWORLD, an Android environment which allows benchmarking of mobile automation agents on a variety of tasks and apps. Unlike prior benchmarks, AndroidWorld dynamically generates novel tasks to create millions of potential unique challenges to measure agent performance on.
Developing chemical synthesis pathways is a difficult challenge for LLMs, as it requires detailed information pulled from external databases and complex, multi-step reasoning. The authors propose ChemCrow, an agent which outperforms GPT-4 alone at planning difficult synthesis pathways for novel molecules.
The authors determine that a single metric derived from an LLM’s performance on a variety of common benchmark tasks is highly predictive of its abilities on more complex agentic challenges such as those found in AgentBench. This corresponds with the observation that more powerful LLMs such as GPT-4 can succeed at agentic tasks that are beyond the capability of earlier and weaker models.
FinRobot is a tool which decomposes financial questions into logical problems and uses a suite of agents based on both base foundation models and fine-tuned financial LLMs to generate accurate responses.
Machine translation can translate texts between languages with high accuracy, but the results often fail to capture the flair and contextual nuance of the source material when applied to literature. TransAgents is a multi-agent virtual company which simulates the human translation process, achieving superior results on human evaluations of literary quality.
🛠️ Useful stuff
Webinar on how to use OpenDevin, a community-driven effort to build an open-source version of Devin.
CopilotKit allows developers to integrate AI agents powered by LangGraph with interactive app frontends.
Cover-Agent is a tool for automated creation and validation of unit tests, inspired by Meta’s TestGen.
memary allows LLM agents to store information in knowledge graphs for later retrieval, acting as long-term memory.
đź’ˇ Analysis
An interview with Chi Wang, creator of Microsoft’s multi-agent framework AutoGen. Wang discusses the past, present, and future of both AutoGen and the broader LLM agent ecosystem.
The author speculates that financial advising is the next industry ripe for disruption by agent systems. Algorithmic trading already dominates the stock market, but the more human skill of advising clients on investments may be seeing its moment with the introduction of systems like GPT-4o and Astra.