Introduction

If you're reading this, chances are you’ve encountered the idea of AI agents or are curious about how artificial intelligence pushes boundaries in practical, advanced ways. Over the past few months, I’ve been delving into the capabilities and limitations of agentic AI technology. In the process, I’ve encountered numerous academic papers and technical documents that are dense and challenging to interpret. My aim with this Primer is to simplify these complex concepts and make them accessible to a broader audience. The document is structured into six sections as follows:

1. Evolution of Artificial Intelligence

2. Introduction of AI Agents

3. Enterprise Adoption Patterns

4. Future Trends to Watch

5. Key Players in AI Agents

6. Investments & M&A

While this Primer emphasizes agentic AI, it is not exhaustive of the broader field of artificial intelligence. If there is enough interest, topics such as foundational models, generative AI, energy and compute innovations may be explored in future articles. 

I hope you enjoy it and please reach out if you want to discuss more!

(Thanks to Kitan Ademidun for reading draft and suggestions)

Let's dive in.

Evolution of Artificial Intelligence

The groundwork of AI began in 1950 when Alan Turing published “Computer Machinery and Intelligence” which proposed a test of machine intelligence called The Imitation Game, which has continued to develop in the last few decades as computers and data become more accessible.

In recent years, breakthroughs like AlphaGo (2015), Transformers (2017), and ChatGPT (2022) have brought AI to the hands of everyday people.

• There has been a 5x increase in sentences mentioning AI one year after the release of ChatGPT across earning calls for S&P500 companies

• Global AI spending, including applications, infrastructure, and services, is projected to double to $632 billion by 2028, driven by a 29% annual growth rate fueled by the rapid adoption of generative AI

• Two-thirds of YC companies in S24 were in AI (149 out of 235 companies), 70% in applications, 28% in toolings, and 2% in infrastructure/foundational models 

The application of AI is multi-faceted, meaning it can be used in a wide variety of fields and situations, from healthcare diagnostics to self-driving cars, with the potential to revolutionize different industries and aspects of daily life

The evolution of AI can be broadly described into three stages:

• Predictive AI: Predictive AI is the stage where technology leverages machine learning to provide enormous-scale predictions across use cases at scale. It ranges from your text autocorrect functions to playlists and show recommendations.

  • Examples include Algorithmic Feeds,  Recommendation Engines, Auto Suggest

• Generative AI: Generative AI is the stage where AI can be prompted to generate text, code, images, and videos, based on specific prompts or inputs. This technology leverages advanced machine learning models, to generate outputs that are not simply copies of existing data but new and original creations

  • Examples include ChatGPT, Claude, Perplexity, Adobe Firefly, MidJourney

• Agentic AI: Agents represent the next step in the evolution of AI applications and the deployment of machine learning. They are compound systems of AI models with advanced reasoning abilities and access to software tools they use to make decisions and perform complex tasks. They combine the ability of AI models to understand and respond more closely to how people communicate with the efficiency and reliability of software and can do so at scale. 

  • Examples include Salesforce Agentforce, Microsoft Copilot Studio, Hebbia, Voiceflow.

AI Agents are set to transform a new wave of innovation, accelerating digitization and expanding how computers can augment human capabilities. Research from OpenAI and the University of Pennsylvania found that with access to a large language model (LLM), workers could complete about 15% of tasks faster without sacrificing quality. Adding software tools built on top of LLMs, like Vertical SaaS applications, jumps to between 47% and 56% of tasks. 

For example, Klarna CEO Sebastian Siemiatkowski plans to leverage AI Agents and reduce its workforce by 50% by the end of 2025. Salesforce CEO Marc Benioff also announced that Salesforce will not be hiring software engineers in 2025.

Introduction of AI Agents

An AI agent refers to a system or program capable of autonomously performing tasks on behalf of a user or another system by designing its workflow and utilizing available tools.

Unlike general-purpose technology such as ChatGPTs and Perplexity, AI Agents are designed with a much more narrowed scope to meet a particular objective, such as addressing customer inquiries or helping doctors with patient record documentation.

Every AI agent follows a 3-step process to perform autonomous actions 

1. Perceive: AI agents gather and decode information from sources like sensors, databases, and interfaces to turn data into insights. They pinpoint meaningful patterns and extract what’s most relevant in their environment.

2. Reason: A large language model (LLM) guides the reasoning process — understanding tasks, crafting solutions, and coordinating specialized models for jobs like content generation or image analysis.

3. Act: Agents perform tasks by connecting with external systems through APIs. Built-in guardrails ensure safety and compliance — like limiting insurance claim processing to specific amounts before human review.

Then inside every agent, three critical components are essential to help agents perform

• Tools: Mechanisms through which AI interacts with the external world, enabling tasks such as accessing APIs, using plugins, and integrating with systems

• Memory: The ability for AI to retain context and learn from past interactions, ensuring continuity in multi-step processes and improving adaptability based on historical data

• Knowledge: A blend of pre-trained expertise and real-time information, allowing AI to reason effectively and navigate diverse scenarios, ensuring accurate insights and precise task execution

The snippet below represents an AI agent in its simplest form. An agent leverages a language model to solve problems, and its configuration can vary widely. The definition of your agent may consist of a large or small language model to use, memory, storage, external knowledge source, instructions, etc.

For example, modern agents like Windsurf and Cursor enable users to prompt, run, edit, build, and deploy full-stack web applications within minutes. These tools support code generation and app development using various web technologies and databases.

AI Agent Design Processes

Although AI agents are autonomous in their decision-making processes, they require instructions and environments defined by humans. There are three main influences on autonomous agent behaviour:

• Architects: The team that design and build the agentic AI ecosystem (“stacks”)

  • The goal is to ensure scalability, performance and security while being cost-efficient across its technology stack and data platforms

  • Decision inputs include system solutions, operating/data standards, interaction models, system interfaces, role and permissions, metadata records

• Developers: The team that deploys the agent and provides the user with access to it

  • The goal is to develop AI agents using enterprise toolings and data while adhering to operational standards, to deliver value to end-users, internal or external. 

  • Decision inputs include model selection, safeguards settings, communication channels, access & entitlements, agent fine-tuning, human-in-the-loop processes

• Users: The user interacts with the AI agent with specific goals via prompting

  • The goal is to materialize value from AI agents, whether it is from time savings to quality enhancement to other value-add use cases

  • Decision inputs include conversational prompts, parameter settings

At each layer, agents must be carefully designed and tested by users to minimize biases, and hallucinations and ensure agents will involve humans in the loop for nuances. The role of a

A human-in-the-loop (HITL) approach is essential for designing AI agentic systems. Instead of relying entirely on machine algorithms, humans can provide real-time feedback on the agent’s actions, and in some cases, step in to guide or override the process. This collaboration ensures the system remains aligned with human goals and adapts to nuanced, unpredictable situations.

AI agents are best designed with a flow diagram because it helps visualize the process steps sequentially. It requires an in-depth level of detail about the actions, decisions, data consumed, and triggers for escalation and resolutions.

For example, when creating an agent to handle appointment booking, users have to configure and align the actions and data inputs.

• What is the opening conversation like in a phone call? Can you personalize it to your customers by including what you know about them? (e.g., name/locations?)

• How do we perceive and validate responses by users? (e.g., making sure the appointment is within business hours and not mistake time ranges as appointment lengths)

• Where can you get technician availability information for the assignment? How to ensure work is fairly distributed among workers and does not conflict with their vacation time?

For more complex agents like customer interaction chatbots, dozens of use cases will become their flows. Continuously anticipating, testing, and iterating is critical to ensure AI agents can meet users’ evolving needs. The more precise and exhaustive the flows users develop their AI agents with, the less room for mistakes they can have.

AI Agent Architecture

The construct of AI Agents also varies depending on the capabilities and use cases. However, most advanced AI agent operations are built on top of a multi-agent orchestration architecture, because when you build a single AI agent to handle all types of use cases

• Coding gets complicated and becomes difficult to maintain

• Agent gets confused with inconsistency responses due to its expanded role across use cases

• Agent gets slower and requires frontier models to respond to various user needs

A multi-agent architectural framework is developed to decouple various capabilities and give each agent a clear role and objective. The modular approach has enabled enterprises to assemble and dissect agents with relevant toolings, knowledge bases and instructions.

To illustrate this with an example, a complex task like writing software, a multi-agent approach would break down the task into subtasks to be executed by different roles — such as a software engineer, product manager, designer, QA engineer, and so on — and have different agents accomplish different subtasks. Mimicking the world before AI where managers have to delegate work to individuals with the right skills and knowledge.

Though the same LLMs power these AI agents and operate within the same environment, decomposing capabilities into multiple specific roles allows for

1. Specialization: While modern LLMs can process vast inputs, breaking tasks into subtasks allows agents to concentrate on specific goals, improving their effectiveness. 

2. Modularity: This design pattern mirrors real-world project management, where specialists divide tasks. Developers can adopt this abstraction to simplify and optimize complex workflows, akin to decomposing a program into processes 

3. Control: By explicitly defining how agents communicate, you can direct their interactions more precisely, avoiding the unpredictability of implicit function calls. This control ensures that agents collaborate in a structured, goal-oriented way, enhancing the overall process.

Emerging frameworks like AutoGen, Crew AI, and LangGraph, provide rich ways to build multi-agent solutions to problems. MetaGPT has also developed advanced frameworks by encoding Standardized Operating Procedures (SOPs) into prompts, enabling error reduction and result verification.

Enterprise Adoption Patterns

As AI spreads, companies are racing to lead in using it to transform how they operate. Across industries, businesses are pouring resources into experimenting with AI, hoping to outpace their competitors and seize an edge.

Enterprise AI software spending has already jumped to $4.6B in 2024 (up from $600M in 2023) – and this is just the beginning as our “service as software” paradigm takes hold.

Incumbents invent new waves of AI products and services 

• JP Morgan has piloted COiN (Contract Intelligence), its unassisted virtual agent that performs manual repetitive tasks such as interpreting loan applications. COiN can save approximately 360,000 hours of review for the firm's legal teams

• Salesforce revealed Agentforce as its next-generation application for enterprises which is going to be the focus of the company in the future

• Walmart deployed AI-powered negotiation software by Pactum with a text-based interface to connect with vendors and suppliers 

Code engineering, support chatbots, and enterprise search + retrieval have become the most popular enterprise use cases. As a result, companies such as Github, ServiceNow, and Box have all transformed into AI-native applications.

Code generation stands out as a key pioneer use case as AI has reached the level of precision and reasoning humans have when it comes to coding. In Oct 2024, Google’s CEO Sundar Pichi revealed that 25% of the company’s code is written by AI, depicting the significance of AI in Google’s operations. 

According to a study conducted at Microsoft, Accenture and an anonymous Fortune 100 electronics manufacturing company with AI toolings available for 4,800+ software developers. The result suggested a 26.08% increase in completed tasks among developers using the AI tool. Notably, less experienced developers showed higher adoption rates and greater productivity gains.

The productivity gains from AI are not evenly distributed across product teams. AI may improve engineering productivity by 25%+ but its effect may not be the same for designers, product managers, etc. Organizations must reconfigure themselves to address imbalance (e.g., PM-to-engineer ratio) to deliver optimal quality in the environment with agentic AI technology. 

Regarding Agentic AI, Capgemini Research Institute suggested that over 10% of organizations are already adopting AI agents in their organizations, with 53% planning to introduce AI agents in their workflows in 2025. 

Though no specific data reveals the adoption patterns for AI agents, the trajectory is clear: AI agents will fundamentally reshape how organizations function. As these agents take on increasingly complex tasks, they won't just complement human work—they'll begin to substitute it.

Consider Salesforce, a titan of enterprise software. It generates $35 billion in annual revenue, an impressive figure constrained by incremental growth. Now contrast that with the $1.1 trillion global companies spend annually on sales and marketing salaries. The opportunity for Agentic AI disruption is many multiples larger than Salesforce could ever be.

This represents ~120 million workers and almost $2.6 trillion worth of salaries. Certainly, AI will not be replacing the entirety of it, but AI will begin to substitute some parts of it. 

The rise of AI agents and applications rests on a critical dependency: the falling cost of the foundational value stack, from semiconductors to infrastructure to LLMs. This follows a familiar pattern in technology—Moore’s Law applied to AI. The deflationary force of technological progress has driven the cost of LLM inference down at an extraordinary pace.

In November 2021, GPT-3 was the only model capable of achieving an MMLU of 42, but at a steep cost: $60 per million tokens. Fast forward three years, and Llama 3.2 3B, offered by Together.ai, achieves the same benchmark for just $0.06 per million tokens. That’s a 1,000x reduction in cost in less than three years. 

The rapid decline in LLM inference costs enables the development of AI-native applications, including AI agents, by improving unit economics. This opens new opportunities for startups and enterprises to deploy AI applications across all sectors. The reduced cost fosters innovation, drives competition and accelerates adoption for enterprises. 

AI Development Trends 

In this section, I have highlighted some of the pivotal development trends poised to define the trajectory of AI and AI Agents in the coming years. 

• Commoditization of General Purpose LLMs Inference Tokens

  • The declining cost of token cost and prices as the model continues to scale. In an unregulated environment, a “race-to-the-bottom” phenomenon will occur until the LLM providers offer major differentiation points or a consolidation of market players. They will now have to compete on volume (distribution)
    

• AI-led productivity gains may cannibalize the traditional B2B SaaS pricing models

  • Legacy companies like Salesforce and Microsoft charge for their software services based on the number of users or seats within a client organization. If AI features allow clients to handle increased demand with fewer users, these companies would need to revise their pricing model. For example, Intercom is charging $0.99 per ticket resolution
    

• AI Agent Market Competitive Landscape Between Incumbents and Startups

  • There are two classes of AI agent providers “incubments” that offer systems of record solutions (i.e., Salesforce, ServiceNow) and “startups” that focus primarily on AI-native agentic solutions (i.e., Harvey, Hebbia). There are no proven records or patterns on how they perform against each other yet, however, 2025 seems to be a year where there will be a lot of competition for enterprise buyers

• AI Transformation Remains as an overlooked gold mine:

  • Accenture, a global technology consultancy, recorded $1.2B in GenAI booking revenue from Sep - Nov 2024, equivalent to $4.8B annually. The figure is higher than the creator of ChatGPT, OpenAI which recorded a $4B revenue in 2024

  • For enterprises to transform themselves with new AI capabilities, they need much more than software or AI solutions, that includes strategy planning, change management, and large-scale transformation for AI tools to be implemented, socialized and adopted
    

• AI Voice Agents Applications for B2C Markets

  • Voice agents that are accessible via devices such as smart watches, personal vehicles, and smartphones can also develop new user-AI interaction opportunities. For example e.g., ChatGPT’s Voice and Inflection’s Pi)
    

• Talent Acquisitions between BigTech for AI talents

  • The race for top AI talent is fierce and highly competitive. Key players are willing to invest significant amounts of capital to recruit and retain AI talents from rivalries 

    • OpenAI is setting new benchmarks with unprecedented compensation

    • Microsoft is recruiting executives from AI startups like Inflection and deepening its partnership with OpenAI through major investments

    • Amazon has made significant investments in Anthropic

    • Google has acquired Character.ai 

    • And more…… 

M&A & Investments

In 2024, AI investments have amounted to $55 billion across its value chain from applications to the energy layer, representing more than one-third of total venture funding. Most of its findings have been invested into the foundational layer, the largest rounds include:

• OpenAI (Feb 2024, $6.6B at $157B Valuation, Led by Thrive Capital)

• xAI (May 2024, $5B at $50B Valuation, Participated by Sequoia, a16z, QIA)

• Anthropic ($4B at undisclosed valuation, Amazon as sole investor) 

• Safe Superintelligence (Sep 2024, $1B at $4B valuation, Participated by a16z, Sequoia) 

• ScaleAI (May 2024, $1B at $14B Valuation, Led by Accel) 

Check out this article by TechCrunch for a list of AI startups that raised over $100M in 2024. 

What is truly different about the AI investment landscape is the participation of strategic investors, including the M7 and infrastructure leaders such as Nvidia and AMD. According to the Financial Times, in 2023, Microsoft, Google and Amazon contributed two-thirds of the $27bn AI investments

Source: The Strategy Deck, Charting how BigTech and Nvidia Invested in Generative AI Startups

Many of these tech giants are crowding out traditional tech investors for the biggest deals in the industry. They offer benefits to startups beyond capital including

• Access to AWS, Azure or GCP computing and storage resources at custom rates

• Distribution for the foundation models through 3rd-party development platforms, such as Amazon’s Bedrock, Google’s Vertex AI Model Garden and Microsoft’s Azure Model Catalog

• PR and joint marketing initiatives targeted at AI application developers and enthusiasts

As a result, most of these leading incumbents have become shareholders of the most prominent AI startups, with a few trajectories in mind 

1. Acquisitions: Acquire the talents and products to vertically (or horizontally) integrate with their existing products and services, to defend and protect their market dominance

2. Infrastructure Partnership: Supply compute and infrastructure services to these startups and convert them into long-term customers 

3. Competitive Alliance: Drive board decisions that align with company strategy to better integrate with your products, or prevent competitors from integrating with its products

Notable Players & Startups in AI Agents 

If you are interested in learning more about active startups involved in AI Agents, I’d recommend you to check out the market maps published by Felicis Venture, Foundational Capital and Insight Partners)

Source Used:

1. https://academic.oup.com/mind/article/LIX/236/433/986238

2. https://www.stlouisfed.org/on-the-economy/2024/sep/ai-optimism-uncertainty-what-can-earnings-calls-tell-us

3. https://www.idc.com/getdoc.jsp?containerId=prUS52530724

4. https://kelvinmu.substack.com/p/yc-ai-companies-a-detailed-breakdown

5. https://arxiv.org/abs/2303.10130

6. https://www.cnbc.com/2024/11/13/buy-now-pay-later-giant-klarna-files-for-us-ipo.html

7. https://www.salesforceben.com/salesforce-will-hire-no-more-software-engineers-in-2025-says-marc-benioff/

8. https://www.salesforce.com/blog/what-are-large-language-models/

9. https://www.salesforce.com/agentforce/what-is-agentic-ai/

10. https://www.youtube.com/watch?v=7pvEYLW1yZw&ab_channel=AWSEvents

11. https://codeium.com/windsurf

12. https://arxiv.org/pdf/2308.00352

13. https://langchain-ai.github.io/langgraph/concepts/multi_agent/

14. https://www.cnbc.com/2024/07/25/techs-splurge-on-ai-chips-has-meta-alphabet-tesla-in-arms-race.html

15. https://www.chasealum.org/article.html?aid=1692

16. https://www.salesforce.com/agentforce/

17. https://hbr.org/2022/11/how-walmart-automated-supplier-negotiations

18. https://menlovc.com/2024-the-state-of-generative-ai-in-the-enterprise/#765bf53e-c1df-477f-941d-810743936402

19. https://a16z.com/llmflation-llm-inference-cost/

20. https://www.youtube.com/watch?v=7pvEYLW1yZw&ab_channel=AWSEvents

21. https://finance.yahoo.com/news/over-25-google-code-now-151413292.html

22. https://hbr.org/2022/11/how-walmart-automated-supplier-negotiations

23. https://www.capgemini.com/us-en/insights/research-library/generative-ai-in-organizations-2024/

24. https://www.linkedin.com/feed/update/urn:li:activity:7269741422210154500/

25. https://substack.com/@kylepoyar/note/c-71224890

26. https://open.spotify.com/episode/2n6EjfBnGczYd1x7YPWaei?si=a3091d5aee324d43

27. https://investor.accenture.com/~/media/Files/A/Accenture-IR-V3/quarterly-earnings/2025/q1fy25/accentures-first-quarter-fiscal-2025-results.pdf

28. https://www.theinformation.com/articles/openai-projections-imply-losses-tripling-to-14-billion-in-2026

29. https://a16z.com/ai-voice-agents/

30. https://techcrunch.com/2024/10/02/openai-raises-6-6b-and-is-now-valued-at-157b/

31. https://techcrunch.com/2024/11/22/anthropic-raises-an-additional-4b-from-amazon-makes-aws-its-primary-cloud-partner/

32. https://techcrunch.com/2024/09/04/ilya-sutskevers-startup-safe-super-intelligence-raises-1b/

33. https://scale.com/blog/scale-ai-series-f

34. https://www.ft.com/content/c6b47d24-b435-4f41-b197-2d826cce9532

35. https://a16z.com/llmflation-llm-inference-cost/

36. https://arxiv.org/pdf/2308.00352

37. https://sensortower.com/blog/state-of-ai-apps-2024

38. https://www.capgemini.com/us-en/insights/research-library/generative-ai-in-organizations-2024/

39. https://sema4.ai/blog/the-five-levels-of-agentic-automation/

40. https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4945566

41. https://arxiv.org/pdf/2303.10130

42. https://arxiv.org/pdf/2309.07864

43. https://www.felicis.com/insight/the-agent-economy

44. https://foundationcapital.com/ai-service-as-software/

45. https://www.insightpartners.com/ideas/ai-agents-disrupting-automation/