Agents that heal themselves. Alexander De Ridder on resilient AI

Alexander De Ridder on Building the Internet of Agents

Episode Summary

Alexander De Ridder, Co‑Founder and CTO of SmythOS, breaks down the coming Internet of Agents. He explains how SmythOS, an MIT‑licensed open source Agent Operating System, separates secrets, runtime, and agent logic to deliver safe autonomy at scale. We get into autonomy versus control, multi‑agent collaboration, why RPA chains break, and how founders can deploy real agent teams in production today.

Key Takeaways

• Agents should handle intelligent, repeatable processes. Humans should handle unique, judgment-heavy work.
• SmythOS is an Agent Operating System. It separates secrets, runtime, and logic. It includes a visual debugger and Agent Weaver for natural‑language agent creation.
• The right design point is safe autonomy. Avoid brittle RPA chains and avoid “bot nanny” supervision loops.
• Multi‑agent systems unlock outsized output. Teams of agents can run support backends, interpret market signals, and produce large content volumes with minimal human review.
• The near future is an Internet of Agents. Think ant colony behavior. Individual units network into a superorganism that compounds capability.
• Craft matters. Alexander treats engineering like artisan work. Build systems that are not only functional but durable, elegant, and meaningful.
• Play accelerates learning. Reserve time to push new models to their limits. Creative side quests inform better product decisions.

Practical Use Cases For Founders

• Customer support and operations. Tiered agent handling with escalation to humans only when needed.
• Fintech and data products. Agents call APIs, interpret signals, and take actions across systems.
• Marketing operations. Agent teams can research, draft, edit, and publish at scale with one human in the loop.
• Internal workflows. SOP‑like processes with intelligence can be fully handed off to agents.

Frameworks From The Episode

1. Autonomy–Control Barbell

• Autonomy. Agents act without constant human approval.
• Control. Guardrails, permissions, auditability.
• Design goal. Safe autonomy that minimizes human babysitting.

2. Production‑Grade Agent Architecture

• Separation of concerns. Secrets. Runtime. Agent logic.
• Tool use first. Database calls. APIs. Models. Computer actions.
• Observability. Visual debugging. Tracing. Recovery paths.
• Human‑in‑the‑loop. Escalation when thresholds or uncertainty are exceeded.

3. Repeatable Intelligence Hand‑Off

• Identify processes that require reasoning yet follow a known SOP.
• Encode SOP as agent tasks and tools.
• Add fallback strategies and self‑healing behaviors.
• Measure outcomes. Expand surface area.

4. Differentiation Map

• RPA. Brittle if/then chains that collapse on failure.
• Bot‑nanny agents. Constant human approvals that kill throughput.
• Agent OS. Runtime designed for tool use, resilience, and real autonomy.

Notable Quotes

• “I paint with code and teams. The internet is my canvas.”
• “2023 is the fossil record of AI tool use. From language to tools to swarms.”
• “Agents should work in the background, heal, and only ask humans when they must.”

Resources

• SmythOS
• SmythOS GitHub

Links

• aiforfounders.co
• ryanestes.info