Ultimate AI Agents Masterclass for Founders & Marketers | Rethinking Marketing with AI Agents
Executive Summary
Think School host Ganesh Prasad talks with Kedar, introduced as head of marketing at Kotak Mahindra Bank and formerly at Amazon India, about building multi-agent AI systems for marketing and operations. Kedar explains automation versus agents with a vending machine and a smart employee, demonstrates an agent turning a four-word brief into a credit card creative, and describes the three research layers, the act, reason and memory components, and splitting work into specialised agents under an orchestrator. Coverage comes from approximately the first 50,000 characters of third-party captions of a 63 minute episode, so the customer-support example is only partly covered. Claims about savings, job loss (the intro cites an Anthropic CEO saying half of entry-level jobs will disappear), and agent results are the speakers' statements, not independently verified. This is not employment, financial or legal advice. Captions are automatic with errors. Topics are untimed; metadata follows YouTube. Source: https://youtubetotranscript.com/transcript?v=GtwVwqPiouE
Key Takeaways
Automation versus agents
Kedar says a vending machine is traditional automation with fixed steps, while an AI agent is like a smart employee who takes a goal, makes a plan, tries it, and reworks it until the goal is met.
The five-step loop
He describes an agent that perceives, reasons, acts, evaluates and adapts, and says it stops only when the goal is achieved. The host adds an example of a motion graphics agent that checks its own output against a quality bar.
Demo: a four-word brief
Kedar types a short brief about a credit card and the agent produces an HTML email and a static creative within minutes. He says it chose a zero forex markup message because it evaluates what drives adoption for a segment at a given time.
Three layers of research
The host explains that an agent for a bank is trained on secondary research from public sources, primary research and historical data. He warns that the output may not be reproducible without the same context.
Act, reason and memory
Kedar says an agent needs components for acting, reasoning and memory. Mixing too many roles, such as director plus finance plus scientist, overloads context and increases hallucination. He says agents are just prompts with API connections and memory.
Split into specialised agents
The host sketches a content system with a topic scouter that uses trends and news, an orchestrator that evaluates titles, and other agents. Kedar says dividing scope this way reduces error.
Customer support agent design
They describe giving a support agent the logic you would give a human: what counts as a bug versus a feature request, P1 to P4 priority levels, and decision rules for duplicates. The agent cannot invent this logic itself.
The Conversation Breakdown
Automation versus agents
⏱ Untimed topicKedar says a vending machine is traditional automation with fixed steps, while an AI agent is like a smart employee who takes a goal, makes a plan, tries it, and reworks it until the goal is met.
The five-step loop
⏱ Untimed topicHe describes an agent that perceives, reasons, acts, evaluates and adapts, and says it stops only when the goal is achieved. The host adds an example of a motion graphics agent that checks its own output against a quality bar.
Demo: a four-word brief
⏱ Untimed topicKedar types a short brief about a credit card and the agent produces an HTML email and a static creative within minutes. He says it chose a zero forex markup message because it evaluates what drives adoption for a segment at a given time.
Three layers of research
⏱ Untimed topicThe host explains that an agent for a bank is trained on secondary research from public sources, primary research and historical data. He warns that the output may not be reproducible without the same context.
Act, reason and memory
⏱ Untimed topicKedar says an agent needs components for acting, reasoning and memory. Mixing too many roles, such as director plus finance plus scientist, overloads context and increases hallucination. He says agents are just prompts with API connections and memory.
Split into specialised agents
⏱ Untimed topicThe host sketches a content system with a topic scouter that uses trends and news, an orchestrator that evaluates titles, and other agents. Kedar says dividing scope this way reduces error.
Customer support agent design
⏱ Untimed topicThey describe giving a support agent the logic you would give a human: what counts as a bug versus a feature request, P1 to P4 priority levels, and decision rules for duplicates. The agent cannot invent this logic itself.
Core Frameworks & Ideas
Automation versus agent - Kedar's analogy
A vending machine follows fixed steps; a smart employee pursues a goal and adapts.
Perceive, reason, act, evaluate, adapt - Kedar's loop
An agent reads context, plans, executes, checks the result and tries again.
Specialised agents under an orchestrator - host's design
Narrow agents plus a coordinator reduce the scope for error.
Actionable Next Steps
- ✓Write a one-page brief for a task as if for a smart new hire, with the goal and the quality bar.
- ✓Split one workflow into two or three narrow roles instead of a single do-everything prompt.
- ✓List the rules a human would need for your support inbox and write them down for an agent.
- ✓Test any agent output against the bar you set before using it.
Notable Quotes
“A vending machine is a traditional automation. A smart employee is an AI agent.”
“agents are just prompts”
“you've reduced the scope of error”
The One Thing to Remember
An AI agent is a goal-driven worker, so brief it with context, constraints and a bar for quality like you would a new hire.