The Shift to Agentic AI – Reuven Cohen (Talk Outline)

A Craft 2026 morning keynote-style solo talk by Reuven Cohen, a self-described “agentic engineer” for almost four years (a title that once drew “the craziest thing I’ve ever heard of”). The presentation itself was built in Lovable. He frames his ~40-year career — from an 8-year-old with a software-less Tandy to today — as “coming full circle”: the modern agent console looks almost exactly like the Unix systems of the 1980s. The talk threads chronologically: (1) a career of “build and share” (shareware → cyber designer → Napster → Infrastructure-as-a-Service → CloudCamp → US government / Microsoft / EY), (2) the compression of the build cycle across computing waves, (3) agentic engineering, harnesses, and agent swarms, (4) the “dream cycle” and GitHub as a social network, (5) infusing AI into the physical world with contrastive vector intelligence (RuView / RuVector), (6) determinism vs. probabilism and sublinear loops, (7) the human as the highest-order model, (8) the shift to sovereign, cloudless AI, (9) the Agentics Foundation, and (10) closing reflections on reinvention (and his company Cognitum).

1. Framing: An Agentic Engineer Who Came Full Circle

1.1 The presentation built in Lovable

• He opens noting the whole presentation was built in Lovable — “hopefully it’ll work correctly.”

1.2 “Agentic engineer” for ~4 years

• He has called himself an agentic engineer for almost four years.
• Early reaction to the title: “that is the craziest thing I’ve ever heard of.”

1.3 Two apps, 40 years apart

• On screen: the app he built at age 10 in GW-BASIC, and the app he’s best known for now.
• That app was formerly Cloudflow, now Roo Flow — a harness for building agentic systems around Claude Code.

1.4 Starting young gives perspective

• Building since age 8 gives perspective on what can be built and how change affects the industry.
• He feels he’s been “preparing for this moment” his whole life and saw opportunities earlier than most.

1.5 Full circle to the 1980s console

• Modern agent work in a console looks “almost exactly” like what he built in 1987.
• We’ve gone from graphical/Windows UIs and web UIs back to an environment that looks like a 1980s Unix system.

2. A Career of “Build and Share”

2.1 The software-less Tandy (1987)

• His parents bought him his first computer, a Tandy (~1987) — with no software.
• Forced into a paradigm: he had to invent things to run on it.

2.2 Discovering shareware

• He learned others were sharing applications.
• Via his local magazine (PC World), he filled out the back and shared his own code to receive theirs — this was shareware.
• To get software, you had to give software — a formative introduction to computing.

2.3 A democratized, open culture before open source

• The sharing culture democratized access to technology — “free and open” before open source became the predominant delivery mechanism.

2.4 Cyber designer (1994)

• In 1994 he chose a different path than his high-school peers: to be a “cyber designer.”
• That meant building dynamic applications before the world had figured out what the internet was.
• He constantly had to evolve, moving from BASIC to HTML and other environments.

2.5 Napster interface (1999)

• He was an early designer of the Napster interface in 1999 — “a whole fiasco as well.”

2.6 Infrastructure-as-a-Service (2004) and hypervisors

• Around 2004 he saw the opportunity in hypervisors (then emerging) to apply a dynamic, utility approach to infrastructure.
• This gave him the “Infrastructure-as-a-Service” moniker — years before “cloud” was a thing.

2.7 Open-sourcing on SourceForge → AWS calls

• He open-sourced it on SourceForge (the “early predecessor to GitHub”) and it blew up.
• His open-source project of the month was noticed by AWS, who reached out to integrate it into their system.
• Lesson: putting software openly into the ecosystem makes opportunities call you back.

2.8 CloudCamp: 300+ cities

• As an early open-source/cloud advocate, he co-founded CloudCamp.
• It grew to more than 300 cities globally and introduced concepts, people, community, and like-minded development processes.

2.9 Working with giant organizations and the US government

• Large companies (including AWS) saw his software and asked him to help implement it.
• He got a call from the US administration asking him to help with their website — a door-opener (he’s Canadian).

2.10 The template

• The through-line: put software openly into the ecosystem, and opportunities come back to you.

3. The Compression of the Build Cycle

3.1 CTO of Microsoft exploring AI integration

• He describes being CTO of Microsoft at the time, exploring how Microsoft could integrate AI into their ecosystems.

3.2 Traditional programming is long and structured

• Traditional programming: define a PRD and structures, assemble a team to implement — “a long, drawn-out experience.”
• Enterprise applications traditionally took months if not years.

3.3 The shift he saw

• He saw a shift in the way you program: the ability to contract the time to build each piece.

3.4 ChatGPT / Claude Code compress build time

• Something like ChatGPT or, more recently, Claude Code compresses the time to build systems.

3.5 Magnifying the individual

• The compression magnifies an individual to the output of what would have taken a larger team.

3.6 Each computing wave compresses further

• Personal computing — build applications for yourself.
• The web — share with a larger pool of people online.
• Mobile — carry the application into the greater world.
• AI — go from idea to implementation almost instantly (“a game changer”).

3.7 Four years of experimenting and instant sharing

• Over the last four years he’s focused on building the things he imagines and sharing them almost instantly with a global pool of users, contributors, and companies.

3.8 The $1.4B EY AI infrastructure project

• He built one of the first and largest AI-world infrastructures for Ernst & Young — a $1.4 billion project.
• It showed how this could be implemented in large organizations and allowed automation of processes — “just type it in” and make it available quickly.

3.9 From “a tool you call” to “a tool that calls you back”

• The shift: from a tool you call, to a tool that calls you back — giving answers applicable to the specific problem being solved.

3.10 Pushback and the disruptor’s role

• Going into any organization (EY included), the status quo treats someone like him as very disruptive.
• He had to both articulate the opportunity and teach them how to build these applications.

4. Agentic Engineering & the Harness

4.1 What “agency” means

• Agency = a system’s ability to work with little to no human oversight or operational context.
• I.e., a system that can be driven and operated autonomously.

4.2 OpenAI VIP program and Codex (~early 2022)

• Building autonomous systems around early 2022, he worked with OpenAI as part of their VIP program.
• He saw the opportunity to use the then-Codex platform in his workflow.

4.3 Shifting to an automated harness

• He moved from manual implementation to an automated harness wrapped around the ways he traditionally built applications.

4.4 What the harness does

• The harness lets him build, test, and deploy in a fast, iterative pattern.

4.5 The drawback of rapid speed

• Building at rapid speed, he pushed code as fast as possible.
• Two things happened: many projects faded away, while some picked up steam.
• He learned from community interaction (GitHub and others) what worked and what didn’t.

4.6 50 million lines of code

• Over ~four years he’s put out more than 50 million lines of code — “an astronomical amount.”

4.7 “Understanding without seeing”

• No single human can understand/interact with that much code.
• So he built systems to understand without seeing — now a critical part of his applications.

4.8 Claude Flow → autonomous swarms 24/7

• His more popular platform, originally Claude Flow, provides a framework for autonomous development.
• 24/7, multiple agents work as a swarm.

4.9 How the swarm differs from a human team

• The swarm works like a team, but with a critical difference: agents can communicate, collaborate, and contribute to one another’s development process.
• He can adapt it as he builds; the system seems to speed itself up — an increasing velocity of both code development and empowerment to build the unexpected.

5. The “Dream Cycle” & GitHub as a Social Network

5.1 The background “dream cycle”

• Every morning he has a “dream cycle” running in the background.
• It looks for novel new capabilities by scanning the most recent research.
• It targets applications that have never been done but straddle the line between possible and impossible.

5.2 Iterate and publish

• He iterates on those, develops them, and publishes to the internet.
• This “collects a like-minded group” who share his passion for building, sharing, and discovering.

5.3 The velocity “rocket”

• The result is a velocity increase — a “rocket” of new applications that opened access to companies and business opportunities.

5.4 From solo developer to a company

• He is essentially a solo/independent developer but has recently started a company for these technologies.

5.5 Roo Flow’s traction

• Roo Flow has more than 50,000 stars; on a good week, multiple millions of downloads.

5.6 GitHub as social network, code as content

• GitHub has become a kind of social network; code is the content — “the TikTok of code generation.”
• The key: these applications straddle what was possible and what was never done before.

6. Contrastive Vector Intelligence: Infusing AI Into the Physical World

6.1 RuView — sensing the world from Wi-Fi

• RuView (“half the world sees RUV”) is based on research from Carnegie Mellon University.
• It takes information around us from Wi-Fi signals to detect heartbeats, body pose, and other information “beyond our comprehension.”
• It translates that into something visible and understandable; ~70,000 stars.

6.2 AI beyond the internet

• His opportunity: AI sitting beyond the structure of the internet — “in a world where anyone can vibe anything.”
• Not about building the next SaaS tool, but integrating AI into the everyday fabric of society — light, ephemeral, cost-efficient.

6.3 From zeros and ones to contrast

• 40–50 years of technology has been based on zeros and ones (mostly a one).
• His contrastive approach looks not at the information you see but the information you don’t see.

6.4 The 8-billion-agents problem

• If you created an agent for every person — 8 billion people — even at $1/person/month that’s $8 billion/month — “not going to happen.”

6.5 The opportunity is in the void

• Using a hyperbolic space structure sitting at a confluence between a vector space and a graph overlay, he measures what isn’t happening, not what is.
• “The opportunity is in the void” — the system sees the differences as information changes, with reinforcement forming the basis of the graph.

6.6 Radically lower power — runs on an ESP32

• The approach uses significantly less power and compute and requires no GPU.
• What was computationally and electrically heavy can now run on a nominal device like an ESP32 — sensing unseen Wi-Fi signals and heartbeat fluctuations in a room.

7. The RuVector Stack

7.1 Hyperbolic vectors vs. the Euclidean plane

• At the bottom of the stack: RuVector, built on a hyperbolic vector.
• A traditional vector works on a Euclidean plane — “like a piece of paper,” focused in the middle and blurred at the edges.
• A hyperbolic vector is a center point expanding in every direction, giving a native hierarchy.

7.2 Graph / hypergraph overlay

• On top he applies a graph structure creating interconnected relationships of all the information.
• Uses a hypergraph (multi-dimensional graph) to relate the information.

7.3 Graph neural network for attention/reasoning

• On top of that, a graph neural network provides an attention mechanism — what matters, and “how can I reason upon my own reasoning.”
• It does not require a monolithic AI; it’s a lightweight graph playing between the vector mathematics.

8. Determinism vs. Probabilism

8.1 The problem with probabilistic language models

• Most language models are based on probabilism — a GPU running dense mathematics to give the one most-probable answer.

8.2 Some domains need absolutes, not probabilities

• Many things don’t require probability; they require absolutes.
• For a medical/math system, you don’t want the most-probable answer.
• You want a system that can either say “I don’t know” or give an answer via a specific, reproducible series of steps.

8.3 Adapting over time, not pre-trained

• His structure adapts over time rather than being a moment of pre-training (as most models are).
• The structure adapts in parallel to the world it sits within.

8.4 Sublinear-time loops (< 1 ms)

• Orchestration runs in very tight loops in sublinear time — agents running in sub-one-millisecond increments.

8.5 Two quirks of very short time frames

1. Less energy — the shorter something exists, the less energy it uses.
2. Less entropy — the shorter something exists, the less noise/chaos; enabling signal separation from noise.

8.6 Why looping matters in agentic systems

• Agentic systems loop to ask “is it right? is it wrong? should I move on?”
• These loops needn’t run for hours/weeks — they can be ephemeral, tightly constrained moments.
• Each loop understands its own single-problem context, then communicates to other loops/agents that adapt and learn quickly.

8.7 Separating signal from noise in the room

• Example: Wi-Fi signals can separate all the noise in the room — e.g., separate the cell phone in your pocket from your heartbeat and breathing pattern.
• Slight permutations as the Wi-Fi signal returns from transmitter to processor are rendered as a graph.

9. The Human as the Highest-Order Model

9.1 The graph is the critical part

• The graph enables information to be separated, understood, and acted upon quickly.

9.2 Model choice is secondary to escalation

• In his structure, the highest-order model is a human.
• Choosing a Claude model, Codex, an o-series, or Kimi K2 is secondary to the ability to escalate.

9.3 What the system decides to escalate

• The system decides whether to send to another agent, send to another person, or fetch additional context from a tool/resource it has access to.

10. Sovereignty: Cloudless, Owned AI

10.1 From cloud/online guy to sovereign

• He started in the online world — ran a BBS as a kid, involved in early online tools.
• Over the last couple of years he shifted to a sovereign environment: AI that doesn’t need the cloud, sitting in an owned, fully controlled environment.

10.2 The future balance

• The evolution likely balances a small group of very large infrastructure providers (the Amazons) against a new opportunity: building anything on your own technical stack, independent of a network.
• The stack is moving from traditional approaches to ones owned and controlled by an individual.

10.3 The Agentics Foundation

• ~two years ago, a group created the Agentics Foundation, focused on adoption of agentics and on individuals who can create and operate independently of any central organization.
• It grew quickly to dozens of cities; they just had a packed event in Budapest.
• It’s a learning-together community exploring opportunities and challenges.

11. Closing Reflections

11.1 It’s not about 50 million lines of code

• The real lesson isn’t the 50M lines — it’s doing things in the least amount of code, in the most practical, cleanest, lightest, most efficient way.

11.2 Constant learning

• His takeaway: he’s constantly learning and applying it — excited each morning to see what contributors and his AI built overnight and how to build “the next big thing.”
• The very act of creating and sharing is what lets him apply applications in novel ways.

11.3 Cognitum

• His company is Cognitum — a device that infuses these technologies into business applications: medical, IoT, various devices that learn in proximity to the world.
• Enables ambient AI infused into “just about everything” without a GPU or cloud.

11.4 Proactive, not reactive

• Traditional technology is reactive; AI lets systems be proactive — listen, understand, and act before something matters, adapting as the world changes.

11.5 Understanding the individual, not everyone

• The technology should understand me as an individual — even a person’s unique electromagnetic signature — rather than everyone/everything.

11.6 Reinvention, not job loss

• Three decades of being “in front of things” (cloud, early AI environments via Roo Flow), a culmination of 40+ years.
• He does not see a “dark moment where we all lose our jobs” — he sees an opportunity for reinvention.
• Agents as an augment to our own capabilities are the exciting inflection point; the opportunities are “far beyond anything I’ve seen in my career.”

People & References Cited

• Reuven Cohen — speaker; self-described “agentic engineer,” ~40-year career; Canadian.
• Companies / orgs: Lovable (built the deck); Microsoft (he cites being its CTO at the time); AWS (found his SourceForge project); Ernst & Young (EY, $1.4B AI infrastructure project); OpenAI (VIP program, Codex); CloudCamp (co-founded, 300+ cities); the US administration (website help); Agentics Foundation (co-created, dozens of cities, Budapest event); Cognitum (his company); Carnegie Mellon University (RuView research inspiration).
• His platforms / projects: GW-BASIC (childhood app); shareware (PC World); Napster interface (1999); early Infrastructure-as-a-Service on SourceForge (2004); Cloudflow → Roo Flow (harness around Claude Code, ~50,000 stars, millions of weekly downloads); Claude Flow; RuView / RUV (~70,000 stars, Wi-Fi sensing); RuVector (the underlying vector stack).
• Hardware: Tandy (~1987); ESP32 (GPU-free edge device).
• Models mentioned: Claude models, Codex, an “o” series, Kimi K2.
• Concepts: coming full circle to the 1980s console; “build and share” as a career template; compression of the build cycle across computing waves; “a tool you call” → “a tool that calls you back”; agency / autonomous harness; agent swarms (24/7, collaborating); “understanding without seeing”; 50M+ lines of code; the background “dream cycle”; GitHub as social network / code as content; contrastive vector intelligence (“opportunity in the void”); hyperbolic vectors vs. Euclidean plane; hypergraph overlay; graph neural network attention; determinism vs. probabilism; “I don’t know” and reproducible steps; sublinear (<1 ms) loops; energy/entropy reduction and signal separation; the human as the highest-order model; sovereign / cloudless / GPU-free ambient AI; proactive-before-it-matters; reinvention vs. job loss.

Video: https://www.youtube.com/watch?v=DUyccVFjlz0 — Transcript via yt-transcript.sh; outline generated from the transcript.