The agentic web

The web is changing. Bots accounted for 50% of global internet traffic in 2023. Our friends at Theory Ventures recently wrote about a future where this figure climbs to 90%. Let’s explore this idea further. At Felicis, we think browsing the web could become as archaic as renting a DVD, and sooner than most people realize. A future like this deserves exploration, so let’s dive into the first- and second-order implications of an agentic web.

First, it’s worth taking stock of how far we’ve come since the ChatGPT singularity a mere 20 months ago. It’s not an exaggeration to say generative AI is reshaping entire industries. Take software development, for example. Copilot accounted for 40% of GitHub's revenue growth this year and is already a larger business than GitHub was when Microsoft acquired it in 2018. Customer support is another prime example of AI’s real-world impact. After one month in production, Klarna’s AI customer support agent handled two-thirds of customer inquiries, performing the work of 700 full-time agents and driving an estimated $40M in incremental profit to the company this year. 

This is just the tip of the iceberg. Agents expand the scope of what LLMs can do by pairing foundation models with customer-specific data, multi-step reasoning, and the ability to take action on users’ behalf, as Sierra succinctly describes here. With this structure, companies like Decagon and Sierra are pushing Klarna’s agentic support model even further, and agentic coding platforms like Poolside, Cognition, and Factory are doing the same for software development. Even in consumer search, tools like Multion, Perplexity, and the soon-to-be-released SearchGPT are redefining how we access information online. The shift to an agentic web is in full swing.

‍

The rise of AI agents

We’ve been following agents for a while now and recently noticed an inflection in search volume, open-source repositories, job postings, and research papers about AI Agents. We’ve even seen a sharp uptick in ads mentioning AI agents. 

‍

‍

These charts suggest interest in AI agents is rising. To meet this future demand, an ecosystem of specialized agentic infrastructure is coming into focus. 

‍

AI agent scaffolding

For agents to deliver on their promise, we need more than foundation models. When paired with the right supporting infrastructure, models like Llama 3.1 405b and GPT-4o are good enough to enable the first wave of AI agents. We map the emerging ecosystem of infrastructure around foundation models below:

‍

‍

Authentication: Anon provides robust authentication mechanisms to verify the identity of AI agents and manage their access controls. This might enable AI agents to log into websites on your behalf in the future. No more manually booking flights?

Security: Invariant Labs is building the security layer between AI systems and the real world. This is critical for establishing user trust, and “LLM security” is a constant moving target.

Agent Frameworks: Langchain and LlamaIndex are leading frameworks for agent development, opening the door to millions of developers to build agentic apps. F500 companies like Databricks, Instacart, and Adyen are already working with Langchain. 

Agent RAG: LlamaIndex and Unstructured enable AI agents to enhance responses with real-time data and contextually relevant information. RAG is a vital piece of nearly any successful agent deployment.

Multi-Agent Orchestration: Multi-agent systems delegate sub-tasks to specialized agents in a chain of actions. This architecture has proved critical for handling complex queries.  

Runtimes: Modal and Browserbase offer environments and resources to execute tasks with minimal latency and maximum scalability. Browserbase’s headless browser, which gives agents the ability to browse the web, has seen rapid adoption in recent months. 

Routing: Martian routes user queries to different models, reducing the total cost of using LLMs. DSPy automatically optimizes prompts before feeding them to a model, eliciting the best possible response. Both companies, broadly speaking, are in the optimization space. 

Memory: MemGPT creates and maintains memory for AI agents, enabling them to retain context and learn from previous interactions over time. Agents that remember user interactions will create more personalized experiences.

Evals: Weave, a tool created by Weights & Biases, allows developers to monitor, measure, and improve the performance of AI agents through rigorous testing and feedback loops. This category has garnered the clearest PMF signal across all agent infrastructure segments. 

No-Code Agents: Brevian empowers business users to create and deploy AI agents without requiring coding knowledge, bringing custom agents to the masses.

At this point, it’s worth stepping back and asking, “How are agents different from traditional automation?” One founder put it nicely in a recent conversation with Felicis: agents handle edge cases well, iteratively converse with users to achieve desired results, and adapt to evolving interfaces. These capabilities are crucial in creating a seamless user experience where AI agents can perform tasks reliably and at scale. CrewAI’s CEO explored this idea further in an energetic talk at this year’s AI Engineer World’s Fair.

For AI agents to function optimally without superintelligent underlying models, other preconditions must be considered. The best agentic applications we’ve seen are grounded in rules, manuals, or standard operating procedures. In other words, the action space of AI agents should be constrained. Companies like Norm, Altera, and Cleric embody this trend nicely: clear rules prevent AI agents from veering off course. Effective agentic apps tend to also have access to structured, proprietary data. In short, agents need the right context to perform complex tasks. With these conditions met and the above ancillary functions for agents unlocked, all the building blocks are in place to create amazing agentic apps today. Over the next 12 months, momentum will shift decisively towards app development.

‍

Where we see opportunities for agents

We think agent adoption will happen in waves. The first categories involve text-based workflows, like marketing, paralegal, SDR, SRE, medical scribe, and support agents. Next, we’ll see more complex, multi-modal use cases, like architecture, gaming, security, and education. This was one of our motivations for backing an open-source VLM company, which we’ll be announcing soon! Another axis to consider is data privacy. We think agent adoption in regulated industries like healthcare, financial services, and compliance will be slower than in non-regulated areas. One way to turn this headwind into a tailwind is to build an advisory board that lends credibility to your startup. Norm has done this exceptionally well.

Some areas we’re exploring:

1. Architect: 70%+ of the cost of an architecture firm is on-computer work. Multimodal agents will help draft building designs in seconds, automating a significant chunk of manual work for junior architects.
2. Paralegal: Legal agents will handle drafting, discovery, contract management, and more. One attorney told us the startup he works with already does the work of a third of his paralegal staff.
3. Healthcare Admin: Agents can act as virtual scribes, documenting patient interactions and managing administrative tasks, freeing up time for doctors and nurses. One company in this category is slated to 10x revenue this year.
4. Chief of Staff: 60% of worker time is spent on coordination. AI chiefs of staff will manage calendars, prioritize tasks, and handle interactions with other agents. Everyone will have an AI executive assistant in the same way everyone got an email address in the 1990s.
5. Compliance Analyst: Agents in compliance will automatically monitor regulations and ensure that a company’s activities align with the latest legal requirements. Compliance missteps can lead to millions in fines, a problem made more acute by regulatory sludge. AI agents can help.  
6. Financial Analyst: Financial agents will track expenses, generate financial reports, and provide insights to help with budgeting and investment decisions. One stand-out company here is Finster. 
7. Recruiter: Recruiters spend ⅓ of their time sourcing candidates. Similar to SDRs, AI recruiters will sift through resumes, schedule interviews, and handle initial outbounds for recruiters. 
8. Systems Reliability Engineer: Developers spend 30% of their time reviewing vulnerabilities. AI agents will monitor system performance, predict potential failures, and automatically initiate corrective actions to maintain optimal reliability and uptime.

If you’re building a startup focused on any of these verticals—or ones we may have missed—please have your AI agent reach out to james@felicis.com. 

Agents represent the next frontier of Human-Computer Interaction, and the groundwork is being laid today for an agent-dominated web.