# Stop Guessing: Build Robust AI with Layered CoT

Manish Sanwal, NewsCorp | AI Engineer Summit 2025 | 10:16

Source: https://www.youtube.com/watch?v=VTJHR7rQ2KI
Channel: AI Engineer (https://www.youtube.com/@aiDotEngineer). Summarised by AIE Talks.
Page: https://aietalks.com/talks/stop-guessing-build-robust-ai-with-layered-cot
Published: 2025-02-22
Tags: agents, guardrails, multi-agent, structured-outputs

## TL;DR
- Multi-agent systems divide complex work among specialized AI agents that can be updated, checked, and improved independently.
- Standard Chain-of-Thought prompting makes intermediate reasoning visible, but it remains sensitive to prompt wording and has no built-in error correction.
- Layered Chain-of-Thought verifies every generated thought against a knowledge base before allowing the reasoning chain to continue.

## Summary
Manish Sanwal explains how multi-agent systems and Layered Chain-of-Thought prompting can make AI reasoning easier to inspect and correct. Specialized agents divide a complex task into smaller responsibilities, such as detecting pedestrians, reading traffic signals, or selecting a route. Chain-of-Thought prompting then asks an agent to work through a problem step by step. That makes intermediate decisions available for debugging, but standard Chain-of-Thought can still produce different results from small prompt changes, miss important connections, and carry an early mistake through the rest of the chain. Layered Chain-of-Thought adds a verification step after every generated thought. Each step is checked against a structured knowledge base or external database before it influences the next step. Sanwal argues that this process catches errors earlier, improves reproducibility, and makes the final reasoning easier to audit. He also describes how the method fits into multi-agent systems, where specialized agents collaborate through verified intermediate results.

## Key ideas
### Multi-agent systems divide complex work among specialized agents
[00:24](https://www.youtube.com/watch?v=VTJHR7rQ2KI&t=24s)
Sanwal describes a multi-agent system as a collection of specialized AI agents that cooperate on a complex task. In his self-driving-car example, one agent detects pedestrians, another reads traffic signals, and another checks the best route. This modular structure lets each agent focus on a specific responsibility. Sanwal says agents can be tuned, updated, or improved without rebuilding the whole system. If one agent has a problem, other agents may compensate. The result is a system he describes as more flexible, scalable, and tolerant of individual failures than a single monolithic system.

### Chain-of-Thought makes intermediate reasoning visible
[02:04](https://www.youtube.com/watch?v=VTJHR7rQ2KI&t=124s)
Chain-of-Thought prompting asks an AI model to work through a problem step by step instead of jumping directly to a final answer. Sanwal says this exposes the path the model takes toward its conclusion. That visibility helps people understand how the model is tracking the problem. It also creates an opportunity to find an error in an intermediate step and adjust the prompt or process before the final answer is produced. In his framing, the value is not only the answer, but the ability to inspect the sequence that led to it.

### Standard Chain-of-Thought can carry errors through the whole chain
[04:16](https://www.youtube.com/watch?v=VTJHR7rQ2KI&t=256s)
Sanwal outlines several limits of ordinary Chain-of-Thought prompting. Small changes in wording or context can produce very different reasoning, which makes results harder to reproduce. The process has no built-in mechanism for checking each step as it is generated, so an early bad inference can create a cascade of later errors. Correction usually happens only after the inference is complete. He also says that when a problem has many interdependent factors, the model can miss important connections and produce an oversimplified or incomplete conclusion.

### Layered Chain-of-Thought verifies each thought before continuing
[05:57](https://www.youtube.com/watch?v=VTJHR7rQ2KI&t=357s)
Layered Chain-of-Thought adds a verification stage to every part of the reasoning process. First, an AI agent generates an initial thought or hypothesis from the prompt. Before producing the next thought, the system checks that output against a structured knowledge base or external database. Sanwal gives examples such as a fact-checking algorithm, a consistency check using contextual reasoning, or another model that evaluates accuracy. Only after the thought passes verification does it influence the next step. The chain therefore grows through repeated cycles of generation and checking.

### Per-step verification enables earlier self-correction
[07:35](https://www.youtube.com/watch?v=VTJHR7rQ2KI&t=455s)
Sanwal says that checking each reasoning step helps the system catch and correct errors before they spread through the entire chain. He also argues that independent verification makes the overall process less sensitive to small changes in the input, which improves reproducibility. Because the reasoning is broken into discrete steps that can each be checked, the process becomes easier to audit and interpret. The final conclusion is built from intermediate results that have been validated rather than from an unchecked initial assumption.

### Layered reasoning fits multi-agent collaboration
[08:56](https://www.youtube.com/watch?v=VTJHR7rQ2KI&t=536s)
Sanwal says Layered Chain-of-Thought can be implemented with existing large language model tools and integrated into multi-agent systems. Each specialized agent can contribute to a larger reasoning process while its inferences are validated before the system moves forward. In this design, specialization divides the task and layered verification checks the results. Sanwal presents the combination as a way to improve accuracy and reproducibility while keeping the reasoning process more transparent and interpretable.

## Notable quotes
- "True AI isn't about one giant leap of faith, it's built incrementally with every step verified and refined through collaborative effort." (00:00)
- "Chain of Thought is a method that guides AI to think through the problem step by step rather than simply guessing the answers." (02:04)
- "Before moving on, the generated thought is immediately verified." (06:45)
- "The future of AI isn't just about building bigger models but it's about creating systems that are structured, explainable and reliable." (09:40)

## Tools & references mentioned
- Layered Chain-of-Thought prompting
- Chain-of-Thought prompting
- multi-agent systems
- large language models
- knowledge base
- arXiv paper 2501.18645
- NewsCorp

## Who should watch
- You are designing an AI workflow that currently depends on one large model producing an unchecked final answer.
- Your system needs intermediate reasoning that can be inspected, fact-checked, or corrected before later decisions depend on it.
- You are evaluating whether specialized agents and verification steps can make model outputs more reproducible.

## Related talks

- [Creating Agents that Co-Create](https://aietalks.com/talks/creating-agents-that-co-create) (Karina Nguyen, OpenAI, 24:22)
- [Grounded Reasoning Systems for Cloud Architecture](https://aietalks.com/talks/grounded-reasoning-systems-for-cloud-architecture) (Iman Makaremi, Cat.io, 26:06)
- [How to Build AI Agents that Actually Work](https://aietalks.com/talks/how-to-build-ai-agents-that-actually-work) (Patrick Dougherty, Rosco, 17:44)
- [Design Patterns for AI Trust: Juries, Libraries, and Agent Tiers](https://aietalks.com/talks/design-patterns-for-ai-trust-juries-libraries-and-agent-tiers) (Alex Bauer, Upside.tech, 17:09)
- [Stop AI Agent Hallucinations: 5 Techniques + Production Patterns](https://aietalks.com/talks/stop-ai-agent-hallucinations-5-techniques-production-patterns) (Elizabeth Fuentes, AWS, 55:19)