[3Qs with AI CoE] Guest Fernando Lucas Pérez
Why Single-Agent AI is Legacy Tech: The Case for “Implicit Orchestration”
Most podcasts are 45 minutes of fluff wrapping 5 minutes of insight. We skip the weather, the “how are yous,” and the generic bios.
Welcome to 3 Questions.
The rules are simple: I ask the guest three rigorous, tailored questions based on their actual work. They ask me zero to three questions back. We are done in 15 minutes.
For this episode, I sat down with Fernando Lucas Pérez. Fernando is the VP of AI Automation Solutions at Trilogy, leading the development of Cu Chulainn, a multi-agent platform designed to replace rigid workflows with autonomous swarms.
Here is what we uncovered.
1. The “Implicit” Orchestration (Factory Line vs. Navy SEALs)
I asked Fernando about the architecture of his swarms. Most developers use Explicit Orchestration - a rigid “Graph” where Agent A must talk to Agent B, who must talk to Agent C. This mimics traditional code logic.
Fernando argues this is a trap. He uses Implicit Orchestration.
He appoints a “Manager Agent” within the swarm. This Manager isn’t a hard-coded script; it’s an agent with a “persona” of authority. It listens to the other agents and decides the next step dynamically.
Spiky Point of View:
“Explicit Orchestration” strangles agentic potential.
When you hard-code the interaction path (Agent A -> Agent B), you are treating agents like functions. But agents are probabilistic.
Implicit Orchestration - where a Manager Agent dynamically decides who speaks next - is the only way to handle complex, non-linear tasks. It shifts the paradigm from “following a recipe” to “managing a team.”
2. The “Andon Cord” (Why Swarms are Safer than Agents)
One of the biggest fears with autonomous agents is the “loop of death” or cascading hallucinations. Fernando’s solution is architectural, not prompt-based.
He implements an “Andon Cord Agent”.
Inspired by the Toyota Production System, this agent has one role: Quality Control. It monitors the conversation of the swarm. If it detects a hallucination, a loop, or a deviation from the goal, it “pulls the cord” and stops the process immediately.
Spiky Point of View:
Single-Agent systems have no “Self-Correction” mechanism. If GPT-4 starts hallucinating, it creates a feedback loop of error.
Multi-Agent systems are inherently safer because you can deploy a dedicated “Pessimist Agent” (The Andon Cord) whose only incentive is to find errors. This adversarial dynamic stops bad results before they reach the user.
3. The Taxonomy Trap (Where AI Fails)
We discussed the limits of autonomy. Fernando was clear: AI is a terrible strategist.
He cited his experience with Ticket Classification. AI is incredible at sorting 1,000 tickets if you give it the categories. But if you give it 1,000 tickets and ask it to invent the categories, it creates garbage. AI cannot define your business logic.
Spiky Point of View:
Classification is a matrix of strategic priorities (e.g., is “Urgency” more important than “Topic”?). AI cannot weigh these invisible values.
Humans must build the Tree (Taxonomy); AI can only sort the Leaves (Classification). If you ask AI to build the tree, you lose the operational context of your business.
The Reverse Card: The “Runaway Effect”
Fernando challenged me on the boundary between AI and Data. We converged on the concept of the Runaway Effect.
If an AI system consumes its own generated data without constraints, quality degrades exponentially.
Fernando’s solution is to treat Data Entry as a Deterministic Act. He doesn’t let the AI “write a report.” He forces the AI to call an API with strict schema validation to log the data.
Spiky Point of View:
Never let an LLM “free-text” data into your system.
You must force the AI to use a Deterministic Tool (like an API or a form) to input data. The constraint of the tool acts as a filter, preventing the “Runaway Effect” of bad data polluting your ecosystem.
Watch the full 18-minute video below: