In my last post, From Apps to Agentic Workflows, I explored how business processes have evolved from spreadsheets to SaaS and now to autonomous AI agents. These agents are not just tools; they reason, make decisions, and take action across workflows.
Since writing that post, I’ve dug deeper into how to design, build, and ship agentic workflows. I’ve studied AI platforms, explored the AI stack, followed the discourse on Twitter, read papers, and examined case studies of systems like Intuit Assist1 and Sierra’s agent platform2 3. This research left me with two big questions:
- What does it take to design and build AI agents effectively?
- How will AI agents impact industries like cybersecurity, compliance, and manufacturing, where I see enormous potential?
Let’s unpack this.
What Are Agents?
Traditional software applications are tools: they help you get the job done, but you still do the work. For example, a SaaS tool might help you schedule a meeting, but you still manually check calendars, adjust times, and send invites.
Agents are different. They:
- Reason and act by analyzing context, planning actions, and autonomously executing tasks.
- Interact with tools by calling APIs, updating systems, and solving problems across platforms.
- Continuously learn to adapt to workflows and improve over time.
Sierra frames agents in three ways3:
1. Personal Agents – drafting emails, booking flights, or summarizing documents.
2. Internal Agents – automating complex internal workflows like software debugging or HR tasks.
3. Customer-Facing Agents – acting as brand representatives to manage orders, troubleshoot problems, or engage customers.
For instance, Sierra’s agents helped WeightWatchers resolve 70% of customer sessions autonomously with a 4.6/5 satisfaction score3.
The Layers Behind Agentic Workflows
To build an agentic workflow, you need to understand the AI stack, which consists of five layers:
1. Semiconductors (The Compute Foundation)
Companies like NVIDIA, AMD, and Intel produce GPUs that power the computational demands of modern AI. Without them, foundation models wouldn’t exist4.
2. Cloud Infrastructure (Scalable Compute)
Providers like AWS, Google Cloud, and Azure democratize access to compute power, enabling startups and enterprises to deploy AI workflows affordably and at scale4.
3. Foundation Models (The Brains)
The foundational layer consists of models like GPT-4 (OpenAI), Claude (Anthropic), LLaMA (Meta), and Gemini (Google). These LLMs are the reasoning engines behind agents.
The Batch newsletter notes that these APIs have made AI development dramatically more accessible, with prototyping costs as low as a few dollars4.
4. Orchestration Layer (The Agent Builders)
This is where agents are designed and built. Tools like:
- LangChain and LangGraph help developers chain LLM calls4.
- CrewAI simplifies building multi-agent workflows4.
- Sierra’s Agent SDK provides an OS-like platform to build branded, conversational agents3.
Switching between orchestration tools is complex, so developers need to be deliberate in choosing platforms4.
5. Applications (The Workflows)
At the top sit the AI applications—where agentic workflows come to life. Examples include:
- Intuit Assist for tax preparation: It reads documents, checks for errors, and proactively surfaces tax-saving insights1.
- Sierra-powered agents: Automating customer support for brands like OluKai during Black Friday3.
Now that we’ve explored the components of the AI stack, let’s look at a real-world example of these layers in action.
Case Study: Agents for Highly Regulated Industries
Intuit Assist provides a perfect example of how agents can thrive in regulated industries like finance:
- Automatically reading tax documents for compliance.
- Ensuring accuracy and completeness of filings.
- Escalating complex cases to human tax experts when needed1.
This got me thinking: What if we could do the same for cybersecurity and privacy compliance?
Imagine an agent OS purpose-built for security teams:
1. Compliance Automation: Proactively flagging access control violations and fixing misconfigurations.
2. Incident Response: Detecting anomalies and automating first-level remediation.
3. Audit Management: Generating audit-ready reports for GDPR, SOC 2, or HIPAA compliance.
These agents wouldn’t just reduce errors—they would transform how teams operate in highly sensitive, regulated environments.
Expanded Use Cases of AI Agents Across Industries
While regulated industries like cybersecurity and finance are ripe for disruption, agentic workflows have broader applications:
- Manufacturing:
- Predictive maintenance by monitoring IoT devices to anticipate equipment failures.
-
Quality control using computer vision systems to identify defects in real time.
-
Retail:
- Inventory management with AI agents tracking stock levels and automating reorders.
-
Customer support through personal agents providing real-time order updates and troubleshooting.
-
Healthcare:
- Patient monitoring by analyzing vitals and alerting clinicians to anomalies.
- Automated scheduling of appointments across departments.
Who Builds These Agents?
Building agents doesn’t require ML expertise. Instead, it demands a new role: the AI Engineer. As Latent Space points out, AI Engineers bridge the gap between foundation models and production-ready workflows5.
Core Skills for AI Engineers:
- LLM Application: Using GPT-4, Claude, or specialized models effectively.
- Prompt Engineering: Designing prompts to get consistent, high-quality results.
- Orchestration: Connecting models, APIs, and tools with frameworks like LangChain5.
- Productionization: Turning prototypes into scalable, reliable systems5.
Much like DevOps emerged to unify software and operations, I believe AI Engineers will become essential for deploying and monitoring AI agents at scale5.
Future Vision: Agent OS Platforms
Sierra’s Agent OS concept—enabling businesses to build branded AI agents—offers a glimpse into the future. I see an opportunity to extend this to industry-specific platforms:
- Cybersecurity OS for automating compliance, threat detection, and access control.
- Manufacturing OS for optimizing supply chains, reducing downtime, and ensuring quality control.
- Finance OS for performing regulatory analysis and automating reporting workflows.
What’s Next for Agentic Workflows?
The next step is crossing boundaries. AI agents may eventually converge with robotics, where physical and digital workflows blur. Imagine agents:
- Directing robotic arms in manufacturing.
- Automating last-mile logistics with drones.
- Coordinating operations across smart cities.
At the same time, regulation could become a major barrier. Ensuring agents remain compliant with privacy laws, ethical standards, and industry-specific regulations will be critical.
Whether you’re a developer, founder, or team leader, the time to explore and adopt agentic workflows is now. The possibilities are limited only by your imagination—and your willingness to act.
What workflows in your organization are ripe for automation?
How will you build your agents?