Scott Benson

AI Engineer & Founder of APEX Insights. Building agentic AI systems that automate business operations. Expert in LLM orchestration, GCP, and MLOps.

[Scott Benson @skobyn]

Join Eric Schwartz and myself as we help make sense of the chaos.

Chicago agentics meetup
June 25
6:30 PM Central
Virtual

https://t.co/OaN2vJBj7U

[Scott Benson @skobyn]

Agentic AI
Model Context Protocol (MCP)
Retrieval-Augmented Generation (RAG)
Physical AI
Multiagent Systems
Small Language Models (SLMs)
Test-Time Compute
Vector Databases
Embeddings
Generative Engine Optimization (GEO)
AI Orchestration
Mixture of Experts (MoE)
Edge AI
Synthetic Data
AI Hallucination
Grounding
Prompt Injection
Model Alignment
Fine-Tuning
Domain-Specific Language Models (DSLMs)
Context Window
Tokenization
Explainable AI (XAI)
Federated Learning
AI-Native Development
Zero-Shot Learning
Few-Shot Prompting
Transformer Architecture
Self-Supervised Learning
Liquid Neural Networks (LNNs)
Digital Twin
Embodied AI
Spatial Computing
Differential Privacy
Compute
AI Supercomputing
Autonomous Workflows
Generative Adversarial Networks (GANs)
Prompt Engineering
Large Language Model (LLM)
Reasoning Models
LLMOps
Prompt Chaining
Autonomous Agents
Domain-Specific Agentic AI
Governance-First Agentic AI
Vibe Coding
Heterogeneous Compute
Google Cloud / DeepMind
Anthropic
OpenAI
Hugging Face
RuvNet
SiliconFlow
DeepSeek AI
Cursor / Bugbot
Code Rabbit
Elastic / DeductiveAI
Baseten
Odyssey
Pramaana Labs
Vertex AI Replaced by Gemini Enterprise Agent Platform
Claude-Flow (Ruflo) v3.6 Released
Gemini 3.1 Pro Preview
Agentic AI Moves to Production
Smaller, Specialized LLMs Become Default
Kimi K2.6 & GLM-5.1 Lead Open-Source Coding
Rise of Sovereign Cloud
FinOps Emerges as Standard
Agentic Commerce Shifts Shopping
Alternative Hyperscalers Emerge

Too much to learn. Too much to pay attention to. Just too much.

[Scott Benson @skobyn]

Agentic AI
Model Context Protocol (MCP)
Retrieval-Augmented Generation (RAG)
Physical AI
Multiagent Systems
Small Language Models (SLMs)
Test-Time Compute
Vector Databases
Embeddings
Generative Engine Optimization (GEO)
AI Orchestration
Mixture of Experts (MoE)
Edge AI
Synthetic Data
AI Hallucination
Grounding
Prompt Injection
Model Alignment
Fine-Tuning
Domain-Specific Language Models (DSLMs)
Context Window
Tokenization
Explainable AI (XAI)
Federated Learning
AI-Native Development
Zero-Shot Learning
Few-Shot Prompting
Transformer Architecture
Self-Supervised Learning
Liquid Neural Networks (LNNs)
Digital Twin
Embodied AI
Spatial Computing
Differential Privacy
Compute
AI Supercomputing
Autonomous Workflows
Generative Adversarial Networks (GANs)
Prompt Engineering
Large Language Model (LLM)
Reasoning Models
LLMOps
Prompt Chaining
Autonomous Agents
Domain-Specific Agentic AI
Governance-First Agentic AI
Vibe Coding
Heterogeneous Compute
Google Cloud / DeepMind
Anthropic
OpenAI
Hugging Face
RuvNet
SiliconFlow
DeepSeek AI
Cursor / Bugbot
Code Rabbit
Elastic / DeductiveAI
Baseten
Odyssey
Pramaana Labs
Vertex AI Replaced by Gemini Enterprise Agent Platform
Claude-Flow (Ruflo) v3.6 Released
Gemini 3.1 Pro Preview
Agentic AI Moves to Production
Smaller, Specialized LLMs Become Default
Kimi K2.6 & GLM-5.1 Lead Open-Source Coding
Rise of Sovereign Cloud
FinOps Emerges as Standard
Agentic Commerce Shifts Shopping
Alternative Hyperscalers Emerge

Too much to learn. Too much to pay attention to. Just too much.

[Scott Benson @skobyn]

Agentic AI
Model Context Protocol (MCP)
Retrieval-Augmented Generation (RAG)
Physical AI
Multiagent Systems
Small Language Models (SLMs)
Test-Time Compute
Vector Databases
Embeddings
Generative Engine Optimization (GEO)
AI Orchestration
Mixture of Experts (MoE)
Edge AI
Synthetic Data
AI Hallucination
Grounding
Prompt Injection
Model Alignment
Fine-Tuning
Domain-Specific Language Models (DSLMs)
Context Window
Tokenization
Explainable AI (XAI)
Federated Learning
AI-Native Development
Zero-Shot Learning
Few-Shot Prompting
Transformer Architecture
Self-Supervised Learning
Liquid Neural Networks (LNNs)
Digital Twin
Embodied AI
Spatial Computing
Differential Privacy
Compute
AI Supercomputing
Autonomous Workflows
Generative Adversarial Networks (GANs)
Prompt Engineering
Large Language Model (LLM)
Reasoning Models
LLMOps
Prompt Chaining
Autonomous Agents
Domain-Specific Agentic AI
Governance-First Agentic AI
Vibe Coding
Heterogeneous Compute
Google Cloud / DeepMind
Anthropic
OpenAI
Hugging Face
RuvNet
SiliconFlow
DeepSeek AI
Cursor / Bugbot
Code Rabbit
Elastic / DeductiveAI
Baseten
Odyssey
Pramaana Labs
Vertex AI Replaced by Gemini Enterprise Agent Platform
Claude-Flow (Ruflo) v3.6 Released
Gemini 3.1 Pro Preview
Agentic AI Moves to Production
Smaller, Specialized LLMs Become Default
Kimi K2.6 & GLM-5.1 Lead Open-Source Coding
Rise of Sovereign Cloud
FinOps Emerges as Standard
Agentic Commerce Shifts Shopping
Alternative Hyperscalers Emerge

Too much to learn. Too much to pay attention to. Just too much.

[Scott Benson @skobyn]

@alliekmiller My personal favorite is number 4. I've used that many times. Sometimes it's too honest..

[Scott Benson @skobyn]

@alliekmiller now imagine this as a role based assistant.

GM: "Remind the assistant managers that Friday is trivia night"

Assistant managers agent to assistant managers: "Hey, Dan wanted me to remind you that Friday is trivia night. You should ensure extra staff that night, here's a forecast from last time"

[Scott Benson @skobyn]

I wonder when memory harnesses become the next thing for agents.

We get blasted with information from every angle, all day. Stuff we catch on the news, hear on the radio, pick up from coworkers. Emails, phone calls, zoom, google meets, rss feeds, audiobooks, texts, facebook messages, dms, linkedin articles.

On and on and on.

But on the agentic side, agents get their context from one source. Us.

Yeah, I know there are connectors. mcp's, cli's, the whole stack. But how many people actually configure those the optimal way? And if you run multiple llm's, how painstaking is it to set that up for each one and keep it current? Then once it's configured, how often does the llm actually call it?

What I imagine is one service that pulls your life info into a single place. And all it does is curate.

Been told the same thing 8 times? That might be important. Heard something once, 3 years ago, from a source I didn't know? File that away as not worth my time. Email from an unknown sender trying to get me into their latest skool session? That goes straight to the land of notgonnadothat. Reminder about a reservation coming up? Keep it a little, then dump it once it's over.

Connect that curated memory to your agent. Or multiple llm's. Or openclaw, hermes, nano claw, agent apex, whatever. And I suppose that agent becomes more useful several times over.
And here's the part I keep coming back to. When you change agents, you take your memory with you.
The agent is swappable. The memory isn't.

[Scott Benson @skobyn]

@MikeHoffmann That belief. That validation that you can. That's where it starts.

[Scott Benson @skobyn]

Toby is the layer everyone skips because he doesn't demo well. The doer gets the applause. The watchdog is why the system still works in month three.

I built this same thing into a personal agent memory system months ago. Called it epistemic monitoring. I'm curious how your handling the threshold to flag. How are you measuring it?

[Scott Benson @skobyn]

I've been using loop methods for a while now and will not go back. Design with detail. Plan with precision. Define good. Define tests. Do ALL of that with the help of AI. Then loop through the plans.

https://t.co/Bkem54JKVB

[Addy Osmani @addyosmani]

https://t.co/hIe0UX7z6T

[Scott Benson @skobyn]

This "We will see AI invent" is completely true an unavoidable at scale. The rate of "new" ideas and products is only going to increase rapidly going forward. And that's a good thing. All the worry for ai taking jobs and unemployment will subside as innovation and adaptation increase.

[Scott Benson @skobyn]

This is exactly what I'm seeing in dev cycles right now.
The loop you're describing is the same shape as agentic workflows, just at org scale. Goals become your system prompt and success criteria. Context is your memory layer. Action is tool use and execution. Decisions is the reflection step that rewrites the next run.
Teams building this well aren't doing anything exotic. They're closing the loop on purpose instead of leaving it open. Most orgs run Goals and Action and stop there. No memory, no reflection, nothing wired back in. Every cycle starts cold.
The plumbing point is the one that sticks with me. I've spent more hours on connective tissue than on actual AI work. Getting judgment out of people's heads and into something a system can act on consistently, that's the hard problem. It doesn't show up in the demos.
The autonomous enterprise isn't a destination. It's what happens when you stop leaving the loop open.

[Scott Benson @skobyn]

What type of agent skills are you building right now?

[Scott Benson @skobyn]

@riyazmd774 @threadreaderapp unroll

[Scott Benson @skobyn]

@ia_william @threadreaderapp unroll

[Scott Benson @skobyn]

Wait? Automotive influencers are a thing? Hey @Rivian @Kia @MercedesBenz @VolvoCarUSA @Hyundai @chevrolet if you're looking for a "normal" family of 6 to give an EV SUV with a 3rd row to, you won't be disappointed. We'll share our daily journey which is guaranteed to entertain!

[Scott Benson @skobyn]

@brettfully Funnel

[Scott Benson @skobyn]

@IAmPascio FREE

[Pascio @IAmPascio]

Notion is a FREE employee.

Most people overlook this.

So I wrote The Notion Beginner Manual to teach you how to use it right.

And for the next 24 hrs, It's FREE of charge!

To get it, simply:

• Like
• Retweet
• Reply "FREE"
• Follow me (so I can DM you)

[Scott Benson @skobyn]

@mhdfaran @threadreaderapp unroll

[Scott Benson @skobyn]

@_heyhash PARA