Get Your Weekly Digest

Receive the best podcast summaries and insights delivered to your inbox every week.

We respect your privacy. Unsubscribe anytime.

TED Talks Daily

Everything you need to know about AI agents | Swami Sivasubramanian

with Swami Sivasubramanian

Published November 4, 2025
View Show Notes

About This Episode

In this TED Talk featured on TED Talks Daily, Swami Sivasubramanian explains what AI agents are, how they differ from chatbots, and why they could be one of the most transformative technology shifts of our time. He outlines three key milestones needed for agents to change how we work: transforming software development, establishing trust through automated reasoning, and enabling non-programmers to build and collaborate with agents. Drawing from his own journey and examples from Amazon and Prime Video, he describes a future where human-agent collaboration lowers barriers to creation and makes powerful tools widely accessible.

Topics Covered

Artificial intelligence Automation Software engineering Tools for thought Innovation strategy Data-driven decision-making Entrepreneurship Product development Simulation theory

Disclaimer: We provide independent summaries of podcasts and are not affiliated with or endorsed in any way by any podcast or creator. All podcast names and content are the property of their respective owners. The views and opinions expressed within the podcasts belong solely to the original hosts and guests and do not reflect the views or positions of Summapod.

Quick Takeaways

  • AI agents are autonomous software systems that can reason, plan, and act toward user-defined goals, going far beyond simple chatbots.
  • For AI agents to be transformative, they must first change how software is built by letting developers focus on what to build instead of low-level implementation choices.
  • Trust in agents can be increased by combining agentic AI with automated reasoning to mathematically verify actions before they are executed.
  • Frameworks and infrastructure for building agents are rapidly simplifying, but the interfaces must evolve so that non-coders and business users can also create agents.
  • Human-agent collaboration can drastically reduce the time and effort required for complex creative tasks, such as video recaps, by splitting work into observation, reasoning, and action phases.

Podcast Notes

Show introduction and talk setup

Host introduces TED Talks Daily and the episode topic

TED Talks Daily aims to bring new ideas to spark curiosity every day[1:57]
• Elise Hu identifies herself as the host and frames the show as a daily source of ideas
Introduction of Swami Sivasubramanian and AI agents[2:12]
• The host poses the question of what happens when software can take initiative on its own
• Swami Sivasubramanian is described as a tech leader who will demystify AI agents
• The talk promises to explain what AI agents are, what they aren't, and how they differ from chatbots

Swami's personal story and optimism about technology

Early life and limited access to technology

Background growing up in rural India without technology[2:34]
• Swami grew up in a rural part of India, not in a city and not in an affluent family
• His home did not have a computer while he was growing up
Scarce access to a shared school computer[2:46]
• His middle school and high school had one shared computer for the entire school
• He got about 10 minutes of access per week, at most twice, so 20 minutes maximum to use a computer
• This scarcity meant every second at the computer was precious and had to count
Motivation to learn programming under tight constraints[3:08]
• He wanted to learn how to program despite having only about 10 minutes per session on the computer
• Because time was so limited, he could not spend all day trying out code and debugging
• He had to act as a "human compiler" by detecting syntax errors ahead of time before getting computer access
Falling in love with problem solving and pursuing education[3:28]
• He fell in love with the problem-solving aspect that came with programming under these constraints
• He went on to study at the College of Engineering, Guindy, which he describes as the top college in his state
• He was the first generation in his family to go to college
Pursuing a PhD and anecdote from thesis defense[3:50]
• He eventually pursued a PhD at Vrije Universiteit (Frey University) in Amsterdam
• At his university, two people had to stand by his side during his thesis defense, in case the defense went long and he needed a break
• He asked his brother, who knew almost nothing about the dissertation, to be one of the stand-ins
• His brother was terrified that Swami might step away as a joke and leave him to answer questions, but Swami did not do that

Joining Amazon and helping build AWS

Family reaction to joining Amazon[4:31]
• About 20 years ago, he got a job at Amazon and called his mother to tell her the news
• His mother thought he was going to waste his PhD by joining what she saw as an internet book company, which is how Amazon was perceived at the time
Building foundational cloud and AI services[4:55]
• At Amazon, he had the opportunity to build technologies that became part of AWS
• He cites DynamoDB, SageMaker, and Bedrock as technologies he helped build, which underpin many modern applications
Reflecting on how limited access shaped his trajectory[5:09]
• Looking back, he attributes his journey to the 10 minutes of access he had to a computer that was not even his
• That brief exposure opened up worlds he never thought possible
• As VP of AWS, he now thinks about how AI agents will transform everything and feels optimistic, paralleling his own experience with early tech access

Defining AI agents and distinguishing them from chatbots

High-level definition of AI agents

Core capabilities of AI agents[5:35]
• AI agents are described as autonomous software systems that leverage AI to reason, plan, and adapt
• They pursue user-defined goals and complete tasks on behalf of humans or other systems
• They can sense and interact with their digital environment, converting high-level objectives into executable steps
• They continually learn and improve their efficiency over time
Current and emerging applications of agents[6:22]
• Agents are already being used in software development, drug discovery, precision agriculture, and other domains
• Their ability to use and manipulate interfaces in digital environments similarly to humans lowers the bar for building applications
• Instead of requiring rigid specifications and complex project breakdowns, agents enable users to state their goals and let the agent figure out the steps

Example contrasting chatbots and agents in a research lab scenario

Chatbot behavior in a lab setting[7:42]
• He describes a researcher who asks an AI to help run experiments for exploring a new protein
• A typical AI chatbot would respond by proposing a list of experiments, such as six experiments to run
• This behavior-just proposing experiments-is labeled as chatbot behavior, not that of an agent
What an AI agent would do differently[7:18]
• Given a goal, an AI agent can plan, write code, and use tools to build the experiment setup
• The agent can synthesize the results, reflect on failures, and look for ways to improve its efficiency
• Work that might take a week or more for a human to research and plan experiments could be done in hours or minutes by agents
• In this scenario, the human's role becomes that of a trusted advisor, steering the agent and peer-reviewing its work like a colleague

Lowering barriers to creation and the need for further progress

Impact on traditional constraints like skills and resources[8:25]
• With AI agents, challenges such as lacking a particular skill, insufficient resources, or insufficient headcount for a project begin to fade
• He argues that the future will be shaped by those who can think big and dream even bigger, leveraging agents
Acknowledgment that agents are not yet fully ready[8:29]
• He states that we are not yet at the point where agents can fully deliver on their promise
• He introduces three milestones that agents must achieve before they can fundamentally change how we work and live

Milestone 1: Transforming how software is built

Software pervasiveness and relevance to agents

Digital nature of our world and dependence on software[9:09]
• He notes that much of the world is digital and that in the room, each device likely runs hundreds of applications
• Our daily lives are supported by the work of tens or even hundreds of thousands of software developers
Agents must first reach builders and be useful to them[9:47]
• For agents to reach the masses, they first need to be adopted by builders (developers)
• Developers must find agents useful and interesting for them to survive and spread
• He notes that many tools developers use are already becoming agentic, but ease of building agents is still a key issue

Shifting from implementation details to higher-level architectures

Current burden of implementation choices[10:13]
• Developers currently must make many implementation decisions, such as which server or compute option to use when hosting a site or building an app
• He mentions that if someone has never had to choose how to host a website or application in the cloud, there are many options to consider
• As an example, he describes that one AWS service (EC2) offers about 850 compute options for hosting a mobile app or website
• He emphasizes that this is just one compute option among even more
Agents enabling a refocus on what is being built[11:19]
• In an agentic era, developers can focus more on what they are building rather than how they are building it
• Decisions like which compute instance to choose become less relevant because AI agents will automatically choose those resources
• This shift in conceptualizing effective agent architectures is presented as a key milestone

Milestone 2: Establishing trust in AI agents through automated reasoning

Centrality of trust for agent adoption

Need for trust despite early-stage imperfections[11:31]
• He states that without trust, the capabilities of agents will not matter because they will not be used widely
• He acknowledges that we are in the early days of agentic AI and that agents are imperfect and will make mistakes
• Even for simple tasks, there is an uncompromising need for perfection from users
Agents operate on well-specified systems rather than magic[11:57]
• He points out that agents are not reaching into a magical ether but instead use systems, tools, and environments with well-understood specifications
• Because specifications are well defined, it is possible to mathematically prove whether a system or program obeys its specifications
• He introduces automated reasoning as a technique for providing assurance that a system behaves as expected

Explanation and history of automated reasoning

Definition of automated reasoning[12:25]
• Automated reasoning is described as a field of computer science that seeks to provide assurance about system behavior using sound mathematical logic
• It involves algorithmic search for proofs in mathematical logic
• Automated reasoning can be used to ensure that agentic reasoning is accurate
Historical roots back to Aristotle[12:44]
• He traces the roots of automated reasoning back to ancient Greece
• Aristotle is cited as the first logician to attempt systematic analysis of logical syntax

Case study: Amazon Q and reducing hallucinations

Design goals for Amazon Q as an agent for developers[13:14]
• One of the first agents they built at AWS was called Amazon Q
• Among other things, Q was designed to help software developers build software applications
• The team imagined Q being as smart and capable as their best software developers, able to accelerate roadmaps and obliterate backlogs
Problems in the first prototype: eagerness and errors[13:49]
• The first prototypes behaved like an eager but error-prone intern
• Q hallucinated API calls, generating incorrect or non-existent API requests
• This error-proneness prompted the need for a fix
Using automated reasoning to validate API calls[14:13]
• They formalized all API specifications into mathematical models
• Whenever Q generates an API request, an automated reasoning solver verifies whether it is valid
• If the solver finds an error, it communicates back to the agent, indicating what is wrong and prompting the agent to restructure its code
• This allows many mistakes to be corrected before requiring human intervention

Neurosymbolic feedback loop and performance

Description of the feedback loop[14:41]
• He describes the interaction between the agent and automated reasoning solver as a "neurosymbolic feedback loop"
• The loop is transparent and allows mathematical proof of correctness for an agent's action before the action is taken
Speed and coverage of verification[14:57]
• The verification runs faster than a blink-about 100 microseconds or less for 95% of use cases
• He emphasizes that while this is only a small step, combining agentic AI with automated reasoning will help agents become trustworthy at scale

Milestone 3: Enabling anyone to build and collaborate with agents

Extending beyond developers to all business users

Limitations of focusing only on developers[15:17]
• If efforts stopped at creating a great developer experience, agents would still only serve a small subset of the population
• He notes that in businesses there are many roles, and most people have never written a line of code
Need for interfaces that empower non-coders[16:40]
• The final milestone is enabling anyone-not just programmers-to build agents
• He argues that interfaces for building agents must become familiar to business users

Example: Prime Video recap workflow transformed by agents

Challenge of creating a short recap from long content[15:51]
• He gives an analogy: imagine having only two minutes to recap everything heard in a TED conference, using actual clips from the day
• Simply speaking very fast would not work if the recap must be constructed from selected video clips
• He relates this to Amazon Prime Video, where effective recaps of a series can take weeks to produce and are expensive
• Creating recaps is manual, from designing the story arc to selecting scenes
Agents supporting non-coder experts like cinematographers[16:38]
• Cinematography experts are usually not master coders
• Prime Video introduced agents to streamline the recap process, dividing the workflow into observation, reasoning, and action phases

Three-phase agent workflow: observation, reasoning, action

Phase 1: Observation[16:55]
• In the observation phase, AI agents are tasked with understanding what is happening in the video
• Agents must generate rich, detailed observations about each shot, scene, and the overall story
• These detailed observations are needed to define a coherent story arc and select appropriate scenes for the recap
Phase 2: Reasoning[17:19]
• In the reasoning phase, the agent essentially asks: given what I know, what do I need to do?
• Reasoning layers on top of observation; for example, to create a voiceover narrator for recaps, a reasoning agent generates the script
• The reasoning agent collaborates with the observation agent to ensure the script matches the visual content
Phase 3: Action[17:51]
• In the action phase, trusted human experts work with AI agents to finalize the recap
• This phase blends human judgment with agent-generated content to produce the final two-minute story
Implications for everyday users[17:57]
• He asks listeners to imagine how much easier their own two-minute TED recap would be if they had powerful AI agents following this workflow
• He emphasizes that human-agent collaboration frees humans from drudgery and allows them to create based on what they love

Simplifying agent frameworks and expanding the builder pool

Current state: easier frameworks and agentic cloud infrastructure[18:59]
• Frameworks for building agents are already getting simpler day by day
• Any application developer who can write Python code can now build a pretty useful agent
• He notes that agentic cloud infrastructure is being developed, not just by AWS but across the ecosystem, making it easy to go from proof of concept to production
Need for broader participation beyond programmers[19:11]
• He stresses that simplifying frameworks and infrastructure alone is not enough
• The pool of people who can build AI agents must be expanded significantly
• To achieve this, interfaces for building agents must feel familiar to business users as well as developers

Rethinking how agents are trained and prepared for the real world

Limitations of only making models smarter[19:28]
• He acknowledges that smarter models are valuable but insufficient by themselves
• A world-class scholar who does not take action or ignores how things work in practice is not helpful as an agent
Creating worlds and digital twins for agents[19:43]
• He argues that we need to create worlds in which agents can play and learn
• These will represent the next generation of digital twin environments where agents can practice and improve
• Agents need to be prepared for the real world, not just theoretical tasks

Future outlook and closing challenge

Agents becoming invisible yet transformative

Invisibility of mature agents[20:06]
• He suggests that if all these pieces come together, agents will become invisible in everyday life
• Despite their invisibility, agents will help people do incredible things

Expected impacts in business, science, and entrepreneurship

Acceleration of company creation[20:21]
• In the next few years, he anticipates agents will enable more companies to be created faster than ever before
• He asserts that success will increasingly be determined by one's ideas and ability to describe what they want to build
Breakthroughs in medicine and discovery[20:19]
• He predicts more medical breakthroughs facilitated by agents
• He also foresees many more scientific and other discoveries enabled by agentic systems

Optimism grounded in personal experience and empowerment

Why he is optimistic about an agent-filled future[20:29]
• His optimism is rooted in the belief that the future with agents will ultimately be built by users themselves
• He connects this vision to his own experience of having only 10 minutes of computer access that changed his life
Closing challenge to the audience[20:59]
• He tells the audience that their own "10 minutes"-a pivotal moment of access and opportunity-is coming
• He ends by asking, "What will you build?"

Outro and production credits

Talk context and event information

Event details for the talk[21:00]
• The host notes that the talk was given by Swami Sivasubramanian at TED AI in Vienna, Austria in 2025
Pointer to TED's curation guidelines[20:56]
• Listeners curious about TED's curation can visit TED.com slash curationguidelines

Credits for TED Talks Daily production

Fact-checking and production team[21:03]
• The talk was fact-checked by the TED Research Team
• Production and editing credits go to Martha Estefanos, Oliver Friedman, Brian Green, Lucy Little, and Tansika Sangmarnivong
Audio mixing and additional support[21:16]
• The episode was mixed by Christopher Fasey-Bogan
• Additional support was provided by Emma Taubner and Daniela Balarezo
Host sign-off[21:00]
• Elise Hu says she will be back tomorrow with another idea and thanks listeners for tuning in

Lessons Learned

Actionable insights and wisdom you can apply to your business, career, and personal life.

1

Severe constraints, like limited access to tools or time, can sharpen focus and problem-solving skills, becoming a catalyst for long-term growth and opportunity.

Reflection Questions:

  • • Where in your life are constraints currently frustrating you, and how might they actually help you focus on what matters most?
  • • How could you redesign one of your projects to treat limitations (time, budget, tools) as a way to improve your planning and precision?
  • • What is one area where you could deliberately impose a small constraint this week to force more creative and disciplined thinking?
2

When building complex systems (including AI agents), shifting attention from low-level implementation details to high-level goals unlocks more creativity and leverage.

Reflection Questions:

  • • What projects are you currently micromanaging at the implementation level instead of clearly defining the desired outcomes?
  • • How might your team's productivity change if you specified goals and constraints, then delegated more technical choices to experts or tools?
  • • What is one recurring decision you could automate or standardize so you can spend more time on strategic thinking?
3

Trust in automated systems should be earned through transparency and verifiable behavior, not blind optimism about their intelligence.

Reflection Questions:

  • • Which tools or systems do you rely on today without really understanding how trustworthy they are or how they make decisions?
  • • How could you introduce simple checks, audits, or validations to ensure the outputs you get from software or AI are correct before acting on them?
  • • What is one critical workflow in your work where adding a verification step-even if automated-would significantly reduce risk?
4

Democratizing powerful tools-so non-experts can create and collaborate, not just consume-amplifies innovation across an organization or society.

Reflection Questions:

  • • In your context, which capabilities are currently locked behind specialist skills (like coding or data science) that others could benefit from?
  • • How might you redesign a process, interface, or training so that more colleagues can meaningfully contribute without needing deep technical knowledge?
  • • What is one tool or workflow you use that you could simplify or templatize so others can start building with it themselves?
5

Human-agent collaboration works best when routine observation and synthesis are automated, freeing people to exercise judgment, creativity, and taste.

Reflection Questions:

  • • Which parts of your work are repetitive information-gathering or summarizing that could, in principle, be handled by software?
  • • How would your role change if you spent less time collecting data and more time deciding what to do with it?
  • • What is one specific task this month where you could test a human-tool or human-AI collaboration model, keeping yourself in the advisor or editor seat?
6

Your ability to clearly articulate goals and ideas will become an even bigger advantage as systems emerge that can execute on well-specified intentions.

Reflection Questions:

  • • How precisely can you describe the outcomes you want in your current projects, without jumping straight into how to do them?
  • • In what situations have vague instructions or unclear goals led to misalignment or wasted effort for you or your team?
  • • What is one high-value idea you have that you could write down today as a clear, step-by-step goal description ready for someone-or something-to execute?

Episode Summary - Notes by Tatum

Everything you need to know about AI agents | Swami Sivasubramanian
0:00 0:00