Agents that heal themselves. Alexander De Ridder on resilient AI
with Alexander De Ridder, SmithOS
Agents that heal themselves. Alexander De Ridder on resilient AI
Show Notes
Alexander De Ridder is the co-founder of SmithOS (spelled with a Y: smithos.com), an MIT open-source agent operating system built from the ground up for production AI agents. Alexander came to the United States from Belgium in 2006, built and exited multiple companies, and was building agent infrastructure before “agent” was even a recognized term - fine-tuning GPT-3.5 for tool calling before GPT-4 introduced function calling natively.
SmithOS provides security, separation of secrets and runtime from agent logic, a visual debugger, and Agent Weaver - a built-in interface that lets you describe what you want in plain language and have SmithOS build the agent for you. It is bootstrapped, MIT-licensed, and designed to be a foundation layer for what Alexander calls the internet of agents: the emerging network of AI models collaborating as a super-organism alongside human civilization.
The Problem SmithOS Solves: Between Bot Nanny and Brittle Automation
The current agent landscape sits at two broken extremes. On one end: computer-control agents like OpenAI's Operator, which take over your screen but interrupt you every two seconds asking for approval. The human becomes a bot nanny. On the other end: workflow tools like Zapier and N8N, which give you rigid step-by-step control but break the moment any step in the chain fails. If you've built a long chain of conditionals in N8N and one connection goes down, the entire automation stops.
RPA (Robotic Process Automation) has existed for 15 years - UiPath, Automation Anywhere, IBM Watson all made enormous promises. The reason it never took over the enterprise is the same brittleness problem: if one step in the chain breaks, everything falls apart. Complex work requires intelligence, not just step execution.
SmithOS lives between these extremes. Agents that work autonomously in the background - healing themselves when they hit obstacles, falling back to alternative approaches when a method fails, escalating to a human only when genuinely necessary. Not a bot nanny. Not brittle automation. Controlled autonomy with production-grade infrastructure underneath.
Use Cases for Founders: The VA Onboarding Analogy
Alexander frames agent use cases through the lens of hiring a virtual assistant. If you hired a VA and handed them access to everything with no instructions, they might cause problems. If you onboard them properly - here's the process, here are the systems, here's who to escalate to - they can own entire workflows independently.
Agents work the same way. Any process in your business that requires human-level intelligence but has a known, repeatable step-by-step method can be fully handed off. Where humans are still needed: tasks that are genuinely different every time, where a good SOP can't be written because the situation is always novel.
Real use cases SmithOS customers have built: complete backend support automation with proper human escalation; FinTech solutions with market APIs and AI models interpreting signals; marketing agent teams producing 2,000 articles per week with one human in the loop. Alexander has run enterprise content marketing at scale before - he knows what those team sizes used to look like. The contrast is stark.
The Origin Story: Decimated by ChatGPT, Rebuilt for the Agent Era
In 2019, Alexander launched what he describes as a differentiated AI marketing product - its own semantic embeddings, its own patent, genuinely better results than generic models for internet marketing contexts. When ChatGPT launched, it didn't outperform his product. It outperformed it on ease of use. People didn't want something that took more work, even if the outcome was better. The product was decimated. All immediate competitors went out of business or pivoted.
In the final stages of that company, Alexander's team was building something they called a recipe builder: chained prompts, each step feeding the next. A proto-agent. The pivot question became: what distinguishes a recipe builder from a real agent? The answer: tool calling. Chaining prompts is easy. But when AI models can call databases, APIs, different models, computer interfaces - it becomes a fundamentally different capability class.
When Alexander saw AI models demonstrate real tool use, he saw the entire trajectory of the technology in a flash. That vision became SmithOS, built from scratch for agent infrastructure - not a retrofit of RPA or RAG, but an operating system designed from day one for how agents actually need to run in production.
2023: The Year AI Discovered Tools
Alexander makes a historical argument worth understanding. Human history is the history of tool use. Steve Jobs said the computer is a bicycle for the mind - humans are not the fastest or strongest animals, but give them a tool and they outperform everything. Tool use is the most distinctively human trait there is.
His prediction: archaeologists in the year 3000 will look at the evolutionary history of AI and identify 2023 as the year the first AI fossils with evidence of tool use appeared. What comes next, in order: the era of text and language; the era of tool use (now); the era where tool use becomes sophisticated; the era where agents begin networking and collaborating; the internet of agents as a living super-organism.
SmithOS is designed to be infrastructure for that progression - not just for today's use cases, but for the protocols and architectures that will govern how agents network, collaborate, and form the emerging economy of AI.
The Ant Colony Model: What Networked Agents Actually Become
Alexander's most vivid framework for where agents are going: consider ants. No single ant is particularly intelligent. But networked ants build bridges, wage wars, farm, harvest, form super-colonies spanning continents, and perform surgery on each other - amputating infected limbs to prevent infection from spreading. The individual unit of intelligence is small. The collective achieves extraordinary things.
AI models have vastly more individual intelligence than ants. When they begin networking in the way ants do - when they form the internet of agents - the collective becomes a living super-organism. Alexander draws the parallel to the human body itself: you are a super-organism of millions of cells. Human civilization is a super-organism of billions of humans. The internet of agents is the next layer: a super-civilization of humans and AI working in concert.
Every leap in civilization density - from family to village to city to nation - has been accompanied by enormous technological and economic breakthroughs. The emergence of the agent super-organism will be no different in kind, only in scale and speed.
Painting With Code: The Artisan Philosophy
Alexander grew up in Bruges, Belgium - a culture with a deep artisan tradition inherited from the Romans and carried forward through centuries of Flemish craftsmanship. In that tradition, you don't just open a bakery for economic profit. You open it to become the best bakery in town. The work is not just functional - it is a craft, an expression.
He doesn't distinguish sharply between people and code when he creates. Engineers, agents, technology - these are all tools. The entrepreneur, the creator, uses whatever tools are available to make the vision real. The internet is the canvas: the thing that connects, distributes, and adds the work to the collective fabric of human civilization.
His practical expression of this: every day he plays. New models come out, and rather than just reading about them, he makes a project and pushes it to its limits. This is how he knows where technology actually is - not from specs, but from direct experience of the boundaries. That knowledge is what lets him anticipate where things are going next and ensure SmithOS is building toward that future rather than duplicating what already exists.
Tools & Resources Mentioned
- SmithOS (smithos.com) - MIT open-source agent operating system. Security, secrets separation, visual debugger, Agent Weaver (builds agents for you via natural language). Give them a GitHub star to support the open-source mission.
- Agent Weaver - SmithOS's built-in agent builder. Describe what you want in plain language; Agent Weaver builds the agent for you. Eliminates the need to manually configure agent logic from scratch.
- UiPath / Automation Anywhere / IBM Watson RPA - The 15-year history of Robotic Process Automation that Alexander uses as a cautionary tale. Big promises, limited enterprise adoption, fundamental brittleness of step-by-step chain execution.
- N8N / Zapier - Software 1.0 workflow tools that represent the “full control” end of the automation spectrum. Valuable but brittle: one broken link in the chain stops the whole automation. Not built from the ground up for the agentic future.
Frameworks
The Autonomy-Control Spectrum
Automation tools exist on a spectrum from full control (Zapier, N8N - rigid, brittle chains) to full autonomy (computer-control agents that interrupt constantly for approval). Neither extreme is useful for real production work. The sweet spot: agents that work autonomously in the background, heal themselves when they hit obstacles, fall back to alternative approaches when a method fails, and escalate to humans only when genuinely necessary. SmithOS is built for this middle space.
Tool Use as the Civilizational Threshold
2023 is the year AI demonstrated tool use - the ability to call databases, APIs, and external systems rather than just generating text. Alexander argues this is as significant a threshold as early humans discovering fire or the wheel. The entire subsequent history of AI will be understood as: the text era, the tool use era, the networked agent era, the internet of agents. Founders building now are at the Clovis point - the sharpened flint at the very beginning of that progression.
The Ant Colony Model for Agent Networks
Individual AI models have vastly more intelligence than individual ants. When AI agents begin networking the way ants do - sharing work, specializing, escalating, collaborating - the collective becomes a super-organism. The emergent capabilities of networked agents will exceed what any individual model can produce by orders of magnitude, the same way ant colonies achieve things no individual ant could conceive. The internet of agents is this super-organism forming.
Recipe Builder → Agent: The Tool Calling Leap
Chaining prompts (recipe builders) is easy and was the state of the art in 2022. What transforms a recipe builder into a true agent is tool calling: the ability to reach into databases, APIs, other models, and computer interfaces. That single capability change - from pure text generation to genuine tool use - is what makes agents fundamentally different in kind, not just degree, from prior automation approaches.
Ease of Use as a Moat (and a Threat)
Alexander's 2019 product was technically superior to ChatGPT for its specific use case. It was destroyed anyway - not by a better product, but by a product that was dramatically easier to use. People chose worse outcomes with less friction over better outcomes with more work. The lesson cuts both ways: for founders building AI products, ease of use is often the primary competitive dimension, regardless of underlying quality.
Play as Competitive Intelligence
Alexander plays with new AI models every day - not reading about them, but making projects and pushing them to their limits. This direct boundary-testing is how he maintains a real-time understanding of what technology can and cannot do. From that understanding, he can extrapolate where things are likely to go next and ensure SmithOS is building toward that future rather than the past. Play is how he stays genuinely at the tip of the spear.
FAQ
How is SmithOS different from Zapier, N8N, or OpenAI's Operator?
Zapier and N8N are Software 1.0 - rigid step-by-step chains with AI sprinkled in. If one step breaks, the whole automation stops. They were not built from the ground up for agentic work. OpenAI's Operator is a true agent but interrupts you every two seconds for approval - the human becomes a bot nanny. SmithOS lives between these: agents that work autonomously in the background, heal themselves when they encounter obstacles, fall back to alternative approaches, and only escalate to humans when genuinely needed. It's not an RPA retrofit or a RAG retrofit - it was built as agent infrastructure from day one.
What kinds of businesses benefit most from SmithOS?
Any business with repeatable processes that currently require human intelligence but follow a known step-by-step method. Companies have used SmithOS to: completely automate backend customer support with smart escalation to humans; build FinTech solutions that call market APIs and use AI to interpret signals; run marketing agent teams producing 2,000 articles per week with one human in the loop. The general principle: if you would hire a VA to do it and could write a good SOP for it, an agent can probably do it.
Why did Alexander choose to make SmithOS MIT open source?
The driving philosophy is that the internet of agents should be open for all people. SmithOS is framed explicitly as a gift to humanity - a contribution to the foundation layer of the emerging agent economy, not a proprietary moat. Alexander believes the internet of agents is going to happen regardless. His goal is for it to happen with people - for founders and developers to be participants in shaping it - rather than something that happens to them from closed, proprietary platforms.
What was SmithOS building before GPT-4 introduced function calling?
Before GPT-4 shipped function calling natively, Alexander's team had already fine-tuned GPT-3.5 for tool-calling behavior. They were running agents - in the functional sense of AI models calling external tools and systems - before the term 'agent' meant what it means today. That two-and-a-half year head start on agent infrastructure, built when nobody had vocabulary for it yet, is the foundation SmithOS runs on today.
How does Alexander think about the relationship between humans and AI agents going forward?
He frames it as an expansion of human civilization rather than a replacement. Humans already exist at multiple scales simultaneously: as individuals, families, cities, nations, and as humanity the super-organism. Agents join that stack as a new layer - augmenting and extending human capability rather than replacing it. He explicitly expects transhumanism to enter the picture eventually too. His near-term prediction: for every human in a team, there will be a cluster of AI agents supporting them. Humans become managers and product leads. Engineers who can only code are no longer sufficient - they need to think like customers, marketers, and product people.
What is Agent Weaver and how does it work?
Agent Weaver is SmithOS's built-in natural language interface for building agents. Rather than configuring agent logic manually, you describe in plain language what you want the agent to do, and Agent Weaver builds it for you. It's the answer to the most common friction point Alexander hears: founders who know exactly what they want to automate but struggle with the how. Agent Weaver abstracts the how.
What is Alexander's advice for engineers in the age of AI?
Knowing how to code is no longer sufficient. The engineers who will be most valuable are those who can think like customers, marketers, and product people - who can apply good judgment across disciplines, not just write technically correct code. AI handles increasing amounts of the mechanical execution. Human value concentrates in understanding what should be built and why, and in the judgment required to direct AI systems toward genuinely useful outcomes.