22 February 2024

OpenAI Agents 2.0… Is This The End Of Google Search?

1Frameworks

9Insights

Frameworks in this episode

Marketing4 steps

Monthly LLM Brand Poll

Track how AI models discover, rank, and describe your brand each month.

Read →

Insights & moments

The myth-busts, hot takes, explainers, and tools worth keeping.

Myth Buster· 1

Myth Buster05:00

An Agent Cannot Be Merely Mostly Right

Generating an incorrect answer is different from taking an incorrect action inside a business system. Enterprise agents must address privacy, permissions, and execution reliability because a hallucination could expose data or trigger unintended communications.

Action errors can have consequences beyond a bad text response.
Agents may access sensitive customer and workplace data.
Reliable boundaries and predictable execution are adoption blockers.

“you would not want that bot or agent to hallucinate”
Kieran Flanagan · 05:00

“the actual privacy here and the ability to do things 100% right 100% of the times that to me is going to be one of…”
Kieran Flanagan · 05:30

#privacy#enterprise ai#hallucinations#risk

Hot Take· 4

Hot Take04:00

AI's Next Evolution Is Doing the Work

Chat interfaces can create text and instructions, but agents can carry those outputs into software and execute the task. The hosts see this transition from generation to action as a fundamental change in knowledge work.

Current chat tools often stop after producing information.
Agents can transfer data and operate workplace software.
Task execution substantially expands the value of an assistant.

“it doesn't just like give us information back that it actually will execute work work on our behalf”
Kieran Flanagan · 04:30

“this changes how we do work”
Kieran Flanagan · 05:30

#knowledge work#task execution#automation

Hot Take06:30

The Model Layer May Lose Its Economic Moat

The hosts argue that base language models will become cheaper, more open, and less differentiated. Durable value may instead accrue to proprietary data, specialized use cases, and consumer or enterprise applications built on top of those models.

Base models are expected to commoditize over time.
Fine-tuning with proprietary data can create use-case value.
Agents and applications may become the differentiated layer.
OpenAI has an incentive to build products above its models.

“the data and the use cases and the Agents around them will be what you really have high level of differentiation on”
Host · 06:30

“there is no economic value long term”
Kieran Flanagan · 07:00

#llms#commoditization#ai strategy#moats

Hot Take11:00

Knowledge Workers May Shift From Building to Reviewing

A future worker may supervise queues of tasks completed by agents, approving results and correcting mistakes so the systems improve. This could remove routine software work while concentrating human effort on strategy and creativity, but it also raises concerns about training automation to replace parts of a job.

Agents may perform jobs while humans review their output.
Corrections can fine-tune future agent behavior.
Routine execution may decline as strategy and creativity gain importance.
Workers could help automate significant portions of their own roles.

“there's less building and there's more like reviewing and tuning”
Kieran Flanagan · 12:00

“we are training ourselves out of a job”
Kieran Flanagan · 12:00

#future of work#human oversight#productivity#automation

Hot Take13:00

Search Engines Could Become Task-Execution Engines

Agents can transform search from finding information into completing the underlying objective. Rather than showing links for a salon, flight, or product, an agent may evaluate options and perform the booking or purchase workflow.

Google Duplex demonstrated AI-mediated appointment booking in 2018.
Future search may complete tasks instead of returning links.
Specialized agents could conduct searches on the user's behalf.
Users may interact less directly with conventional search results.

“the search engine naturally becomes a kind of task execution engine”
Kieran Flanagan · 14:30

“they're going to be the Searchers not you”
Host · 14:30

#search#google#task execution#consumer agents

Explainer· 4

Explainer03:00

The Two Types of Agents OpenAI Is Building

The hosts distinguish between workplace agents that learn computer tasks and consumer agents that complete transactions. Both move AI beyond generating information toward executing actions on a user's behalf.

One agent type learns recurring activity on a computer.
Another can search for and complete consumer tasks such as booking flights.
Execution, rather than conversation alone, is presented as the major shift.

“there's one that will sit in the background learn how you do things on the laptop”
Kieran Flanagan · 03:00

“an agent that's going to be able to not just search for things like flights but actually complete those tasks for you”
Kieran Flanagan · 03:30

#openai#ai agents#automation

Explainer09:00

Why Browser Agents Can Automate Software Without APIs

Unlike custom GPT integrations that depend on an application's API, browser agents can learn and reproduce visible user interactions. Extensions may monitor work, create reusable workflows, and control clicks across tools that lack conventional integration points.

Browser extensions are a likely early delivery mechanism.
Agents can learn recurring workflows from observed activity.
Browser control opens automation to software without APIs.
The approach resembles browser-based testing automation.

“it's monitoring what you're doing through the day learning how you do your work”
Kieran Flanagan · 09:30

“this is different because you don't need that app to have an API”
Kieran Flanagan · 12:30

#browser extensions#apis#workflow automation#custom gpts

Explainer15:30

How AI Weakens Google's Traditional Search Moat

Google's advantage grew from ranking signals and behavioral feedback such as links, clicks, and return visits. The episode argues that models capable of interpreting raw internet text may produce strong answers with less clickstream data, reducing the uniqueness of Google's historical feedback loop.

PageRank used linking behavior as a powerful quality signal.
Modern search also learns from clicks and return behavior.
AI can interpret large bodies of text without identical feedback data.
Lower-overhead competitors can challenge established search companies.

“clickstream data and that kind of feedback loop is no longer necessary”
Kieran Flanagan · 16:30

“AI can actually parse and give you the answer that you want”
Kieran Flanagan · 17:00

#google#pagerank#search disruption#perplexity

Explainer19:00

Why Marketers May Struggle to Influence Agent Recommendations

Agents must decide which products to recommend, but their selection logic can be difficult for marketers to inspect. Training data, online brand discussion, legal restrictions, safety policies, and user context may all affect whether a product appears.

Regulated categories can trigger recommendation restrictions.
Online content and brand discussion influence model knowledge.
Marketers may not know why a brand enters or leaves a shortlist.
Discovery monitoring will become harder across multiple models and agents.

“I still don't really understand I don't think we really understand how it recommends one product over another”
Kieran Flanagan · 19:30

“it's going to be pretty difficult to figure out why your brand is not un like the top five”
Kieran Flanagan · 20:00

#ai recommendations#brand discovery#marketing#model policy