Thought Leadership

Agentic AI

We moved from AI assistants to AI agents. Now we're building AI teams. But the real question isn't how many agents to deploy - it's how to keep humans meaningfully in the loop as systems grow more autonomous. 🔹

Featured Essay

AI Assistants Aren't Enough. AI Teams Aren't Enough Either.

Toward a Human-First AI Architecture

For years, the AI story went like this: build an assistant, give it a prompt, get an answer. It was useful. Genuinely. But we were treating AI like a very fast search engine - a tool that responds to queries rather than a participant in work.

Then came autonomous agents. Suddenly, AI could take actions, not just answer questions. It could browse, write, execute code, call APIs. The benchmark obsession shifted to “how long can it run without human intervention?” And this is where I think we took a wrong turn.

The question was never “how autonomous can we make AI?” The question is “how do we design systems where AI and humans are each doing what they do best?”

Humans are irreplaceable when it comes to judgment under ambiguity, ethical reasoning, relationship context, and creative leaps. AI excels at pattern recognition, information synthesis, consistency at scale, and tireless execution. A Human-First AI Architecture isn't about limiting AI - it's about deploying it where it genuinely wins, and keeping humans where they genuinely win.

This is the architecture I'm building toward. And I think it's the architecture that responsible organizations will converge on - not because it's cautious, but because it's effective.

Agentic AIHuman-FirstArchitectureAWS BedrockResponsible AI

Design Principles

Six principles that guide how I think about production agentic AI systems.

01

Human judgment at the center

Agentic AI systems should amplify human decision-making, not replace it. Every architectural choice - from agent boundaries to escalation paths - should be designed around where human judgment adds irreplaceable value.

02

Agents as teammates, not tools

The mental model shift from 'AI as tool' to 'AI as teammate' changes everything: how you design workflows, how you assign tasks, how you measure success, and how you build accountability into the system.

03

Security and governance by default

Production agentic systems require enterprise-grade security from day one - IAM roles, VPC isolation, audit logging, and data residency controls. AWS infrastructure provides the foundation; the architecture determines whether you use it properly.

04

Cost-efficient scaling

Agentic AI can become expensive fast. Thoughtful architecture - right-sizing models for each subtask, caching where appropriate, and building natural stopping points - makes the difference between sustainable and unsustainable systems.

05

Observable and auditable

You cannot trust what you cannot see. Production agentic systems need comprehensive observability: what did each agent do, why, when, and with what data? This isn't optional - it's the foundation of responsible deployment.

06

Built for the real world

Demo-ware and production systems are different species. Real agentic systems handle failure gracefully, have clear recovery paths, and are tested against edge cases that only emerge in the messy reality of enterprise data.

Running Agentic AI on AWS

The AWS services stack for production-grade agentic systems - secure, scalable, and cost-efficient.

A

Amazon Bedrock

Foundation model access & AI Agent Services

The backbone of production agentic AI on AWS - access to leading foundation models with enterprise security, and native Agent Services for building multi-step reasoning workflows.

A

Amazon SageMaker

Custom model training & deployment

When off-the-shelf models need fine-tuning for specific domains, SageMaker provides the infrastructure for responsible model development, evaluation, and deployment at scale.

A

AWS Lambda + Step Functions

Serverless agent orchestration

Event-driven agent coordination at scale - Lambda handles individual agent actions while Step Functions manages complex multi-agent workflows with built-in error handling.

A

Amazon OpenSearch + Kendra

Knowledge retrieval for RAG

Retrieval-Augmented Generation grounded in enterprise knowledge bases. OpenSearch for vector similarity, Kendra for intelligent document retrieval - giving agents access to your organization's actual knowledge.

Building agentic systems? Let's talk. 💡

Whether you're architecting your first agentic workflow or scaling a production multi-agent system on AWS - I'd love to exchange ideas. The future of work is human-AI collaboration, and it starts with the right architecture.

Connect on LinkedIn