13 February 2025

AI Agents in 2025: Where to Start & What Really Works (No Hype)

1Frameworks

9Insights

Frameworks in this episode

Innovation6 steps

Low-Precision Agent Adoption Ladder

Start simple, prove value safely, then expand into higher-impact agent workflows.

Read →

Insights & moments

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

Myth Buster· 1

Myth Buster32:30

2025 Will Bring More Agents, Not Fully Automated Workforces

Joe expects substantial production deployment but rejects the claim that agents will replace whole workforces this year. Complex integrations, legacy desktop software, unreliable end-to-end execution, and high-precision work will keep humans deeply involved.

More human intervention will be needed than hype suggests.
High-precision processes are especially unlikely to run autonomously end to end.
Complex enterprise implementations require significant integration work.
Browsers are attractive because they offer a common interface.
Legacy and desktop systems remain difficult for agents to navigate.

“there's going to be a lot more humans in the loop than a lot of people believe, especially this year.”
Joe Mora · 32:30

“This is not the years like where agents are taking over the like a workforce. That's not it, right?”
Joe Mora · 34:00

#ai hype#human in the loop#enterprise automation#2025

Hot Take· 3

Hot Take04:00

Slow Agents Can Still Win When Latency Does Not Matter

Current browser agents may be dramatically slower than humans, but speed is not always the relevant measure. For asynchronous work, users can launch several agents, switch to something else, and return when the outputs are ready.

Operator was described as slow and clunky.
Background tasks reduce the practical cost of latency.
Parallel agents can create capacity even when each agent is individually slow.
The strongest early use cases are tasks users do not need to watch continuously.

“it's like, 10 times slower than a human doing it, right?”
Kipp Bodnar · 05:00

“I do see a lot of value, especially in kind of, like this idea of you're firing a bunch of them in the background, and…”
Joe Mora · 06:30

#latency#browser agents#automation#productivity

Hot Take11:00

Agents May Deliver Business Value Faster Than the Early Internet

Joe contrasts AI adoption with the internet's early network-dependency problem. A company joining the internet on day zero had little immediate upside because few others were online, whereas companies can implement AI internally and report bottom-line effects within quarters.

Early internet value depended on broad external adoption.
AI can improve internal processes without waiting for universal adoption.
Reported savings are creating executive and board-level mandates.
Aligned executive and employee incentives are accelerating experimentation.

“But if you were online on the internet on day zero, you would have no upside, no impact on your bottom line, because no one…”
Joe Mora · 12:00

“But what we're seeing with AI is companies implementing it in like two quarters later, they're reporting impacts on their bottom line.”
Joe Mora · 12:00

#ai adoption#enterprise ai#business value#automation

Hot Take28:00

CrewAI Rejects Engineering Candidates Who Ignore AI Tools

Joe says CrewAI permits candidates to use any available AI or search tool during engineering interviews because those tools will also be available on the job. Refusing to use them causes an automatic rejection, while the candidate's skill in using them affects the hiring decision.

The interview permits ChatGPT, Anthropic tools, Cursor, and Google.
Tool use is evaluated as part of practical engineering performance.
Not using available AI tools results in an automatic pass on the candidate.
The agents someone builds also reveal initiative and technical capability.

“If they don't use it. It's an automatic pass, and how well they use it actually counts a lot on if we're gonna make them…”
Joe Mora · 29:00

“Well, on the day to day, they're gonna have the access to the shoes anyway, so I want to know how well they use it.”
Joe Mora · 29:30

#hiring#engineering#ai skills#future of work

Explainer· 3

Explainer01:00

What Separates an AI Agent From an Ordinary Chatbot

Joe defines an agent by its ability to receive a task and autonomously navigate toward a result rather than merely continue a conversation. Memory preserves relevant information, while tools let the agent interact with external systems and act on the user's behalf.

Agents receive tasks rather than relying on continuous chat.
Agency means navigating a problem autonomously.
Memory supports persistent agentic behavior.
Tools connect the agent to systems such as CRMs and ERPs.

“You give it a task, and you can leave the room, and then this agent's gonna try to autonomously, kind of, like, true, this idea…”
Joe Mora · 02:00

“I would say those are the two big components. So as these agents are trying to do something, they're going to use those tools to…”
Joe Mora · 03:00

#ai agents#llms#memory#tools

Explainer08:00

Why Users Need to See and Interrupt Autonomous Agents

The hosts and Joe argue that observable execution is an important transitional interface for autonomous systems. Watching an agent work, taking control when necessary, and retaining an audit trail give individuals and enterprises security, reassurance, and operational control.

Visible execution reassures users about autonomous actions.
Takeover controls let humans intervene before damage occurs.
Enterprises need visualization, control, and later auditing.
A drafting agent that unexpectedly sends emails illustrates the risk of unclear boundaries.

“But there's something beautiful about being able to inspect what is happening there.”
Joe Mora · 09:30

“they want to make sure that they can visualize and control this, and they understand what is happening, and they can out it later on.”
Joe Mora · 10:00

#agent ux#human oversight#security#auditability

Explainer20:00

Why Enterprise Agent Programs Are Becoming Centralized

Individual LLM experimentation can expose unapproved information and trap useful knowledge inside isolated teams. Enterprise agent deployments are therefore moving under central technology or AI leadership, which governs models, personal information, reusable use cases, and departmental enablement.

Edge-led experimentation can create data-governance problems.
Useful prompts and practices become siloed when individuals work alone.
Central teams can standardize approved models and PII controls.
Departments can still build tailored workflows within a governed platform.
Technical specialists become essential for custom systems and complex integrations.

“we are talking about AI powered automations, and automation is going to be a very kind of like customized process, because all these companies have…”
Joe Mora · 21:30

#enterprise ai#governance#pii#deployment

Story· 1

Story24:00

How Agents Turn New Sign-Ups Into Personalized Sales Context

Joe describes agents that research a new account, infer how the person and company might use the product, structure those hypotheses as JSON, and write them into HubSpot and the product database. The resulting information powers more relevant marketing and personalized in-product experiences.

The workflow starts when a visitor creates an account.
Agents enrich the person, role, company, and vertical.
They infer three likely individual and company use cases.
Structured findings are pushed to CRM and product systems.
Marketing and onboarding adapt to inferred customer value.

“given the information that they found, come up with hypothesis on how this person is going to use their product.”
Joe Mora · 25:00

“they convert it into JSON, so structured it, and then push it in two places to their HubSpot and into their product database.”
Joe Mora · 25:30

#lead enrichment#sales#personalization#crm#marketing

Tool· 1

Tool30:00

An Agent Workflow That Proposes and Launches Website A/B Tests

Joe outlines a system that studies a company's website, screenshots, HTML, industry, and competitors before proposing conversion hypotheses. The surrounding product lets a human select hypotheses and then implements and measures the chosen A/B tests.

The human supplies a website, industry description, and optimization goal.
Agents inspect the site's copy, screenshots, and HTML.
Competitor research informs proposed conversion changes.
The agents return A/B testing hypotheses for human selection.
The product layer implements and measures approved experiments.

“The input there was, this is my website. This is my description of my industry. This is what I'm trying to optimize. This is the…”
Joe Mora · 31:00

“And then the agents would go around and do everything behind the scenes, and they would come back with all this, like AV hypothesis for…”
Joe Mora · 31:30

#ab testing#conversion#competitive research#website optimization