The Rise of Agentic AI: From Code to Business Value

More and more companies are making the shift to agentic workflows.

The market for agentic AI is estimated to go from about $7 billion this year to around $93 billion in the next couple of years, indicating a massive industry shift.

About 25% of enterprises are already deploying agentic pilots this year. By 2027, that number will jump to 50%, meaning half of major companies will be running some version of agentic workflows.

Unlike traditional automation tools such as Zapier, agentic workflows can incorporate reasoning and adaptability. While actively building or iterating, the agent can catch issues mid-run, adjust its approach, and update its tools, providing a self-healing capability that is very real in that context.

Traditional automation is like building a train track by hand, where you lay every rail and connection yourself. Agentic workflows allow you to instruct a construction crew to build the track, which handles unexpected issues during the build process, resulting in a track built faster with fewer mistakes.

Once an agentic workflow is deployed to the cloud, the self-healing ability goes away, and it behaves more like traditional, predictable automation.

The practicality of these workflows is supported by the fact that LLMs have become reliable enough for production, capable of reasoning, making decisions, and executing multi-step tasks with consistency.

This working memory has a finite capacity, typically around 200,000 tokens; exceeding this limit can lead to 'context rot,' where the model starts losing track of the conversation's middle parts, making up information, or requiring a session reset.

Various elements consume this token space, including system prompts, integrated tools, `.md` files, MCP servers, and the entire conversation history, emphasizing the need for efficient management.

To manage token usage and mitigate context rot, users can employ slash commands such as `/compact` to compress conversation history, `/clear` to reset the session, and `/context` to view a detailed breakdown of token consumption within a project.

This file is read by the AI every time a message is sent, making it crucial to keep it concise to minimize token consumption and improve efficiency.

The content of `claude.md` typically includes three main layers: the 'what' (tech stack, product structure, key packages), the 'why' (purpose of components), and the 'how' (desired operational behavior of Claude).

Claude is capable of generating its own `claude.md` files; users can run the `/init` command to automatically scan their codebase and create a base configuration.

Agentic workflows leverage a 'WAT' framework, standing for Workflows, Agents, and Tools, providing a structured approach to building and managing AI automations.

For projects pushed to public repositories, encrypting files containing sensitive information is a best practice to ensure security.

If an API key carries significant risk, users should run commands involving the key directly in their local terminal, guided by Claude Code, rather than allowing Claude to handle the key directly.

In this mode, the AI agent thoroughly researches and brainstorms solutions to the user's request, providing a detailed plan before executing any actions.

This planning phase significantly reduces error rates and improves the overall quality of the resulting workflow by allowing the agent to consider various approaches.

Plan mode offers an opportunity for human oversight, enabling users to review and approve the agent's proposed strategy before it proceeds with building, ensuring alignment with project requirements.

For instance, when tasked with scraping 622 job opportunities across 21 pages and organizing them into an Excel sheet, the agent in plan mode would first propose a detailed strategy, including tool creation and workflow execution, before initiating the scraping process.

Conversely, Claude Skills are knowledge-based custom instructions or prompt templates that Claude can dynamically load and apply, functioning more like specialized knowledge modules.

Skills enhance Claude's consistency, speed, and overall performance by allowing it to efficiently review available skills and load only the relevant ones based on the user's request, thereby also conserving tokens.

Unlike MCP servers which facilitate interaction with external APIs, Skills serve as internal knowledge or instruction sets that, once installed, are globally accessible across different projects, improving reuse and scalability.

These scheduled tasks are uniquely agentic, meaning they are self-healing and can adapt to errors, unlike traditional deterministic scripts that would simply fail.

For scheduled tasks to run, the Claude desktop application and the user's computer must remain open, as the automation relies on these active components.

A beneficial feature of this system is its ability to handle missed tasks; if the computer is off during a scheduled run, Claude Code will execute the missed tasks once the computer is turned back on.

This automation capability transforms recurring manual tasks, such as a 'morning coffee' skill that plans the day and provides team updates, into automatically executing routines every morning at a specified time like 6:00 a.m.

This model significantly enhances the ability to comprehend nuanced relationships between various media types, allowing for the creation of truly multimodal databases.

With Gemini Embedding 2, the ingestion pipeline for multimodal data is greatly simplified, eliminating the need for manual chunking or complex pipelines that were previously required for tools like Niten.

The model drastically reduces development time, transforming tasks that previously took days into processes completable in under 30 minutes, such as building a chat application for a 68-page PDF instruction manual with embedded images.

Retrieval Augmented Generation (RAG) is a concept where an AI agent, limited by its training data, retrieves external information to enhance the accuracy and completeness of its responses.

In a RAG system, diverse data sources, including documents, videos, and images, are converted into vector points, which are numerical representations of their meaning, using embedding models like Google's new embeddings model 2.

These vector points are then mapped into a multi-dimensional space within a vector database, allowing the AI to query this space and retrieve relevant data chunks—whether text, image, or video—to augment its generation process.

This system establishes a 'source of truth' by enabling the AI to query across different modalities in a shared embedding space, providing more contextual and accurate answers by understanding the relationships between various data types.

The system cannot directly display images or videos but stores text descriptions alongside each embedding, according to Nate Herk.

Querying for multimodal content relies on these text descriptions and embedding coordinates.

Improved metadata and descriptions during the initial ingestion phase significantly enhance the accuracy of search results.

Displaying actual media files requires updating the front-end to serve them inline, moving beyond just providing file names.

Subject matter expertise is crucial for creating precise and descriptive context for non-text media, ensuring the AI understands the content effectively.

Server configuration errors, such as a server configuration error on Vercel, can arise from missing environment variables.

In this specific tutorial, Nate Herk chose to hardcode the webhook URL to simplify the deployment process for demonstration purposes and to avoid configuration steps.

Vercel’s settings provide an option to configure environment variables, which is a more secure practice for production environments to avoid exposing credentials.

Redeploying code via GitHub commits ensures that changes are consistently applied across both test and production versions of an application.

The decision to hardcode the webhook was made to show how changes can be pushed instantly to Vercel without configuring environment variables.

API keys are managed through an .env file to prevent their exposure in conversation histories, ensuring security.

The setup includes creating a 'firecrawl-cheatsheet.mmd' file to provide necessary context to the AI on tool usage and scenarios.

Additionally, a 'claude.md' file is created to define the project's system prompts, specifically instructing Claude Code to focus on data scraping, extraction, screenshot generation, and site crawling.

The project aims to transform Claude Code into a 'second brain' capable of understanding business priorities, team goals, and communication styles.

This involves establishing a structured folder architecture, an evolving 'claude.md' system prompt, and specialized skill files to facilitate the AI's functions.

The user showcases the system by running four parallel agents for tasks like planning the day, conducting team pulse checks, researching new features, and creating social media content.

This approach shifts from conventional chat-based AI interactions towards a more systematized, local project-based framework to enhance productivity and reduce decision fatigue.

Nate Herk describes Claude skills as providing significant 'leverage,' enabling one individual to create processes usable by an entire team, boosting collective productivity.

Skills serve as 'SOPs for AI agents,' automating diverse tasks such as daily planning, project pulse checks, diagramming, and comment analysis, often running in parallel.

The system demonstrated the execution of four different agents simultaneously, completing tasks like morning planning, a project pulse check, an Excalidraw diagram, and YouTube comment analysis in approximately 30 seconds.

The 'skill creator' skill is used within the project to automatically test, measure, and refine these automated workflows, ensuring their effectiveness and accuracy.

Nate Herk intentionally keeps prompts vague to observe how Claude Code plans complex tasks, showcasing its ability to anticipate and structure workflows autonomously.

Claude Code demonstrates proactive planning, generating detailed steps and evaluation processes, even without being prompted for specific technical stacks.

The approach marks a significant shift where non-engineers, such as executives and managers, can focus on defining business objectives and desired metrics.

The AI independently generated a YouTube weekly roundup skill that aimed to analyze videos, comments, views, and engagement from the past seven days, providing a PDF report on insights, strengths, weaknesses, threats, and opportunities.

Agent teams, while powerful, are generally slower and more expensive, making them suitable only for highly complex, multi-area projects requiring diverse specialized agents.

Sub-agents are preferable for sequential processes, tasks that require a single conversation history or context window, and simple tasks where agent teams would be overkill.

Cost control is a critical factor, with Nate Herk recommending limiting agent teams to between two and five members to manage expenses effectively.

Agent teams should be considered when tasks need to be done in parallel, react to each other, assign tasks, and maintain high quality through multiple verification steps.

Nate Herk highlights the transformative capability of automating QA by deploying multiple headless or headed browsers to concurrently test specific features.

This setup establishes a self-validating cycle of testing, identifying bugs, and fixing them without requiring manual intervention.

The AI agent autonomously restarts the server and reruns tests until all system checks are passed, ensuring continuous quality control.

For example, the AI successfully automated a Google search for dentists in California, learning from initial failures (like being blocked by Google) and adapting its strategy to use DuckDuckGo, ultimately finding phone numbers from five different websites.

Nate Herk asserts that businesses primarily prioritize time, money, and focus, and are indifferent to whether solutions are delivered via AI, human assistants, or makeshift methods.

The speaker argues that selling AI agents or workflows is becoming a commoditized and crowded market, leading to a 'race to the bottom' on price.

Successful consulting involves diagnosing specific business problems and offering concrete outcomes, such as reducing customer support workload by 60% or saving 10 hours a week on client onboarding, rather than showcasing complex AI builds.

Business owners value how AI saves them time or money more than the technical details of the implementation, akin to caring about reaching a destination quickly and cheaply rather than the type of vehicle used.

The most practical AI builds, often less flashy, yield the highest return on investment (ROI) compared to elaborate multi-agent projects, contradicting the popularity of complex demos.

Nate Herk advises offering free or very cheap 'paid practice' builds in exchange for testimonials and case studies, which are more valuable long-term than immediate profit.

The speaker suggests offering a money-back guarantee to clients, flipping the risk from the client to the consultant, thereby lowering the barrier for skeptical clients and creating a stress-free learning environment.

The first three to five projects should be viewed as investments in building a portfolio, teaching the process, and gaining confidence, which enables charging higher rates later.

Herk emphasized that his own journey involved projects ranging from $1 to $2,000 up to $30,000 plus, all stemming from initial low-stakes projects that built his confidence and portfolio.