The Artificial Investor - Issue 48: The evolving AI agent landscape and the unmet expectations
✍️ The prologue
My name is Aris Xenofontos and I am an investor at Seaya Ventures. This is the weekly version of the Artificial Investor that comments on the top AI developments of the last few days.
This Week’s Story: OpenAI releases its AI Agent Infrastructure offering
OpenAI released last week the Responses API, which effectively brings its AI agent functionalities to the hands of developers. At the same time. Microsoft announced that it has integrated MCP within its Copilot Studio, the AI agent building platform for developers. This comes a few days after the announcement of a 415 million dollar funding round of Adept, a US-based enterprise AI agent platform and a few months after OpenAI released the leading horizontal agent solution, Operator.
What are the latest developments with AI agents and why are they important? How has the AI agent Tech stack evolved? Why are we still far from meeting the high expectations of 2025 being the year of AI? What’s next in the AI agent race?
Let’s dive in.
🗞️ Latest developments
🧑🤝🧑 OpenAI democratises the Operator
OpenAI released last week the Responses API, which effectively brings its AI agent functionalities (Operator) to the hands of developers. With this API, anyone can build an AI agent that:
can search the internet for specific information, such as news articles, product reviews, or scientific research
can access and process files stored on the internet or on a local computer
can execute code and programs on a remote server
can engage in natural language conversations with users
has security features to protect user data, including authentication, encryption, and access control
The API can be used to create virtual assistants that can answer questions by extracting data from the Web and documents, and run complex analytics on it, while user data is kept safe and secure. Browsing and mouse/keyboard-move capabilities are not at the level of humans yet, but benchmarks show they are best in class.
⚓ Microsoft relies on Anthropic to increase adoption of its AI agent development studio
Microsoft announced last week that it has integrated MCP within its Copilot Studio, the AI agent building platform for developers. MCP will enable Microsoft’s customers to MCP to easily connect to data sources and access the growing library of pre-built, MCP-enabled connectors and tools.
🤔 What is Anthropic MCP?
Hold on. Microsoft has a big partnership with and investment in OpenAI. Anthropic has received large investments from Amazon and Google. How come Microsoft has announced such a strategic integration with Anthropic? Well, let’s start with what MCP is.
Anthropic's MCP is an open-source standard that facilitates the integration of any LLM (not just Anthropic’s) with external data sources and tools. It serves as a universal framework, allowing for seamless communication between AI models and various systems such as databases, file systems, cloud platforms and development environments. In other words, MCP is like the new USB-C port that Apple finally decided to embrace for its phones and it’s universal across different manufacturers and divides from all over the world. We don't have three different charges for our mobile devices at home anymore, the same one works for all of them.
Practically, if a developer of an agent wants to use three different LLMs and three functionalities (e.g. Web search, file search, Google Maps search), with MCP they would have 66% efficiency, as they would need to build three connections (one for each LLM with MCP) instead of nine. MCP comes already with pre-developed connections to many tools, such as Google Drive, Github, Slack, Puppeteer (web scraping), etc.
MCP currently counts more than 1,000 community-built tools (called servers) and a community of more than 1,000 thousand active developers. Publicly-announced use cases include:
Fintech: Block (Square) uses MCP to enable AI agents to access real-time transaction data, customer profiles and payment histories securely.
Sales engagement: Apollo integrates MCP to allow AI agents to connect to CRM systems, email platforms and external APIs to automate lead generation.
Developer tools: Zed, JetBains Cline, Replit and Sourcegraph use MCP to power AI-assisted coding features by connecting to external resources like Git repositories, file systems, and APIs..
LegalTech: Story Protocol uses MCP to automate IP-related tasks by connecting AI to blockchain data, legal documents, and royalty systems, streamlining processes like IP registration, licensing agreement management, and royalty calculation.
Web Scraping: Apify has built an MCP server that lets AI agents access its pre-built scraping tools, allowing agents to pull data from websites without user intervention for tasks like market trend analysis and summarization.
🌍 The AI agent landscape
🧱 The Infrastructure Layer
The infrastructure layer consists of various sub-layers that provide the foundational elements that enable the functioning and orchestration of AI agents.
Services: Applications that carry out specific tasks and/or give access to specific data, such as the various Google applications (Drive, Maps, Email, etc), web scraping applications (Puppeteer, Zenrows, etc), search (Brave Search, Google Search, etc), code repository access (Github, Gitlab), etc. We are used to seeing them as end-user applications, but in the context of agents the roles are flipped on their head and they are an infrastructure sub-layer that powers the agents’s capabilities.
Agentic Frameworks: Frameworks like LangChain, Llamaindex and Microsoft Autogen allow developers to build agents that stitch LLMs with other components. Protocols like Anthropic’s MCP are focused on facilitating connections to a wide range of services.
Agent toolkit platforms: Platforms that use agentic frameworks to provide easy-to-use connection between LLMs and capabilities (i.e. the Services mentioned earlier). The leading platforms are open and utilise MCP (hence why MCP is a big deal). Composio and Arcade.dev are private and focus on a better user experience during agent creation, while Smithery.ai is open source and serves purely as a catalogue of services. There are also some private platforms that provide services within the context of a given application, like Cline’s marketplace, which allows developers that use the AI coding tool to also code agents that use MCP services.
Agentic APIs: APIs offered by AI model providers like OpenAI (Response API) and Anthropic (Compute Use API) that enable developers to create agents with a predefined limited set of capabilities using just a few lines of code.
📱 The Application Layer
The Application Layer consists of applications that enable business users to create AI agents. These tools can be horizontal (domain-agnostic and sector-agnostic) or domain-specific (e.g. lead generation focused) or sector-specific (e.g. hospital patient management focused).
➖ Horizontal players
OpenAI Operator
The most notable horizontal application launched by a large AI company is OpenAI Operator. It can navigate web browsers, click buttons, type text, scroll through pages, and complete multi-step processes—such as booking reservations, shopping online, or filling out forms—without requiring predefined APIs or rigid workflows. It operates in a cloud-based virtual browser, allowing users to watch it execute tasks in real time while retaining the ability to intervene when needed, such as entering sensitive information like payment details. Operator is powered by a model called the Computer-Using Agent (CUA), built on OpenAI’s GPT-4o, which combines vision capabilities (via screenshots) with advanced reasoning to interact with graphical user interfaces. It’s currently available in the U.S. to ChatGPT Pro subscribers ($200/onth) at operator.chatgpt.com. OpenAI has partnered with companies like DoorDash, Instacart, and Etsy to test real-world applications, though it still struggles with complex tasks like creating slideshows or managing calendars.
Microsoft Copilot Studio
Microsoft introduced the ability to create autonomous agents within Copilot Studio, a low-code platform for building custom AI agents. This was highlighted during Microsoft Ignite 2024 (November 2024), where they launched a public preview of fully autonomous agents that can automate complex business processes with minimal human intervention. These agents can connect to organizational data sources, execute workflows and operate across Microsoft 365, Teams, Dynamics 365 and even external platforms. The transition to a paid preview program for these autonomous agents began on February 1, 2025, indicating a formal launch beyond initial testing.
Also, Google announced in December 2024 that it is working on Project Mariner, an experimental AI agent built on the Gemini 2.0 model. The agent is designed to automate tasks within a web browser, specifically Google Chrome. The project explores the future of human-agent interaction by allowing the AI to navigate websites, interact with web elements, and perform actions on behalf of users, all while keeping humans in the loop for supervision. It is limited to a small group of “trusted testers” in the U.S., with a waitlist open for others to join (accessible via the Project Mariner website). Google hasn’t specified when or if Mariner will roll out broadly.
There are a number of startups that have launched their own horizontal agent creation tools:
(Adept): Founded in 2022 in the US, Adept develops enterprise AI agents that automate complex workflows across software applications using natural language commands. They have raised 415 million dollars in total.
(CrewAI): Founded in 2023 in the US, CrewAI offers a framework for orchestrating multi-agent AI systems to automate complex enterprise workflows collaboratively. They have raised 18 million dollars.
(Convergence): Founded in 2023 in the UK, Convergence builds versatile AI agents with long-term memory to reclaim time for users in consumer and enterprise settings. They have raised 12 million dollars.
(MultiOn): Founded in 2023 in the US, MultiOn creates autonomous AI agents that handle online tasks end-to-end, simplifying routines and boosting productivity. They have raised 7 million dollars.
🔍 Domain-specific and sector-specific players
Salesforce is an example of an incumbent who has doubled down on AI agents. Salesforce Agentforce is a suite of autonomous AI agents developed by Salesforce and announced at the Dreamforce 2024 conference, designed to enhance business operations by automating tasks across various functions. Key features include autonomy, multimodal capabilities (web chat, email, SMS, WhatsApp, and Slack, escalating to humans when needed), integration with the Salesforce ecosystem, human reasoning and a low-code agent builder tool. Examples of agents include a Service Agent for resolution of customer issues, a
Sales Development Representative (SDR) agent that qualifies leads, answers questions, and schedules meetings autonomously, a Marketing Agent that builds campaigns, segments audiences and optimizes performance and a Commerce Agent that acts as a personal shopper, offering tailored recommendations on a website. Pricing starts at $2 per conversation and it was rolled out with the full functionality in February 2025. Salesforce aims to deploy one billion agents by the end of 2025.
CB Insights has released an extensive AI agent market map that covers 26 different categories, across domains (cybersecurity, sales, accounting, HR, marketing, customer service, etc) and verticals (Healthcare, Legal, Finance, Industrials, etc.)
🚧 Adoption barriers
Despite the advancements and potential benefits of AI agents, several barriers hinder their widespread adoption:
Closed-loop systems. While OpenAI’s recent agentic API releases are a great advancement in the field, they cannot lead to massive adoption as they oblige developers to use OpenAI’s models, which will not suit everyone due to functionality, single vendor lock-in, price, etc.
Reliability. AI agents are built on LLMs, which by definition hallucinate and make errors. OpenAI Operator is the best-in-class horizontal agent and its performance is substantially worse than that of humans. As a result, agents are limited to simple tasks and require better tools around them to guide them and help them avoid errors. Also, we need better observability and control tools to ensure
Safety. AI agents inevitably handle sensitive data, such as national insurance numbers and credit card details. Users and particularly enterprises need to get comfortable with the security around agents.
Usability. There will always be a tradeoff between functionality/ flexibility and usability across the various applications. However, being still in the early stages of the market, the current usability threshold for the most useful tools and protocols is quite high, requiring specialised development expertise. This limits the size of the current user base.
Scalability. Particularly when considering enterprise use cases, the tools are not apt yet for large-scale implementations. For instance, try scraping and summarising 1,000 websites on a regular basis using OpenAI Operator or its Responses API. You will find yourself looking at load-balancing solutions, caching infrastructures, proxy applications, etc.
🔮 What to expect next
In the next 1-2 years, we expect AI agents to become more multimodal (images, audio and video), improve their reasoning and become able to break down more complex tasks, achieve high levels of personalisation (tailoring responses to individual users based on past interactions and preferences). Also, we expect to see remote server connections that enable enterprise-grade authentication. As tool marketplaces grow (and potentially introduce monetisation for developers), further standardisation will democratise AI, letting smaller teams build powerful agents and automate business processes end to end. Also, it is worth highlighting that a large part of the agent architecture is local first, which we expect to unlock edge use cases, such as IoT.
Future agents could recognize and respond to human emotions, enhancing their utility in areas like customer service or mental health support. Also, we have already seen the first signs of agent collaboration. In the future, agents might work in “swarms,” pooling their capabilities to tackle complex challenges, such as scientific breakthroughs or large-scale coordination in disaster response.
In the long term, it is clear that AI agents will play a key role in the technological landscape and society. As agents take on more responsibility, they may need to make decisions aligned with human values, requiring advanced frameworks for moral reasoning.
The agent market is moving very fast and we are very excited to follow its development through this space.
See you next week for more AI insights.