RESEARCH

How AI is quickly shifting from chatbots to agents

You knew agents were the buzziest topic of 2026, but new data shows that they're dramatically changing the way workers use AI — and not just in terms of productivity and automation. 

On Thursday, OpenAI unfurled new economic research that showed how knowledge workers are adopting AI agents. OpenAI centered the research around its Codex app, since it has such high-fidelity data on usage. Perhaps most interesting was the data OpenAI shared in the paper about the Codex usage patterns inside OpenAI, which could serve as a canary in the coalmine showing how power users across the enterprise will soon be using these kinds of tools.

Ronnie Chatterji, chief economist at OpenAI, told The Deep View, "The paper you have today is trying to analyze the data to tell the world what's happening with AI... We wrote this paper about Codex, and [we're] really opening up for the first time an economic study of agents. We'll later have other papers on different parts of the AI stack that we think are important."

Here are the most consequential takeaways from the report:

• Agents replace chatbots: Inside OpenAI, every department now uses Codex as its primary AI tool, with the average worker now generating 85% of their output tokens (the results from prompts) in Codex rather than ChatGPT. The bottom line is that you can have Codex go query ChatGPT for you, as well as use a lot  of other tools. It can also carry out multiple queries and/or tasks at the same time.

• Units shift from minutes to hours: Of the users sampled in the study, 70% have made at least one Codex request that would take the equivalent of more than one hour of human work, and 25% have made a request that would take the equivalent of 8 hours of human work. The big deal here is that power users are rapidly adapting to the idea of giving agents work and letting them churn on it in the background. 

• Job functions expand: The most profound insight from the study was that it discovered AI empowered workers to take on new tasks and workloads that would have previously been outside their domain of expertise. So it wasn't just AI automating work, but empowering workers to participate in adjacent aspects of work that essentially made them more valuable.

It's not the only signal that employees are gravitating towards agents. Data by KPMG published earlier this week finds a similar trend: Employee adoption of AI agents has already reached 68%, the firm found, as only 2% of leaders report significant pushback on AI deployments.

Hat tip to OpenAI for using its economics team to mine usage data and share insights on the impacts that AI is having on work, employment, and worker adoption of these tools. If you're not familiar, the same team that produced this paper also produces OpenAI Signals, which publishes data, insights, and stories of real-world AI usage. The data released clearly has benefits to OpenAI, as it would like more people to adopt Codex to  stay competitive in its race with Claude Code. It also generates more token usage than ChatGPT, which naturally supports the company's bottom line. One of the ways OpenAI could gain even more credibility and brand trust would be to publish data that doesn't have clear economic benefits for the company, but provides ample benefits to society by helping the public track and understand the latest AI adoption patterns.

TOGETHER WITH KALSHI

The First Regulated Perps Market in the US Is Live

Perpetual futures (perps) are now live on Kalshi. Offshore, perps do ~$90 trillion in volume annually — until now, none of it was legal in the US.

Trade on whether an asset's price goes up or down. No expirations, no rolling costs, no complex payoffs.

Try Kalshi Perps today.

POLICY

Can new coalition save US jobs from AI?

Tech companies are banding together to address the mess their tech is poised to create. 

On Thursday, a bipartisan group of tech companies, led by former governors, Democrat Gina Raimondo and Republican Eric Holcomb, launched a $500 million initiative called Raise Us, aimed at helping the American workforce adapt to AI. 

The backers include OpenAI, Anthropic, Amazon, Microsoft and Google, and will leverage both private and philanthropic capital. 

• The nonprofit will work with these backers to design programs that help them retain and redeploy employees, measuring its success by avoiding layoffs and helping workers land and keep jobs. 

• The organization operates across four core areas: An employer coalition with a "direct line of sight" into how jobs are changing, state partnerships for workforce policy and programming, education and training for an AI-enabled workforce, and a policy lab to adapt and design new regulations. 

• The funds will initially help support pilot programs in Arkansas, Connecticut, Maryland, and Utah to ease the economic disruption for workers who have been impacted by the tech. 

Raimondo, who will serve as CEO of the initiative, said in a statement that if powerful frontier AI systems leave Americans behind, it won't result in our leadership in AI. Rather, we will have "automated our own decline." 

"America has a technology strategy for leading the global AI competition. It does not yet have a people strategy — and we cannot lead without one," said Raimondo. 

In a LinkedIn post, Holcomb called preempting AI's impact an "all-hands-on-deck moment," adding that the initiative "gives state leaders a playbook that connects more Americans with the skills and careers needed in the years ahead.”

The coalition provides yet another harbinger that AI is leading workforces towards upheaval. Study after study has been published attempting to estimate the damage. While some forecast massive disruption, others estimate that the impact won't be nearly as severe as some claim. In either case, major companies are now using AI to justify cutting large swathes of their workforce, including Meta, Amazon, Microsoft, Block, and others.

It's noble that an initiative like this is attempting to get ahead of AI-inflicted damage on jobs and the economy. Participation from major AI firms, too, is a sign that these tech companies are aware of the harm that could result from the tech they are building and promoting. It's something several of these firms have called out before, such as OpenAI launching the Economic Research Exchange to study the economic impacts of the technology, and Anthropic releasing policy proposals aimed at addressing potential disruption. In the immediate term, scaling the deployment of their technologies would only benefit their bottom lines, even if it potentially wipes out jobs. But in the long term, these companies are clearly aware that continuing down that path will only make them the architects of their own downfall. 

TOGETHER WITH DEEL

AI is changing how global teams are managed

AI isn't coming to HR—it's already here.

From predictive hiring to automated compliance checks, global organizations are using AI to make workforce decisions faster and with less risk. But adoption without a framework creates new problems: bias, data exposure, and audit gaps.

Deel's new report breaks down exactly how leading HR teams are deploying AI across borders—and how to do it in a way that scales.

Download guide

HARDWARE

Chips may be the only hope for controlling agents

Agents have repeatedly demonstrated the ability to overcome software constraints. Hardware might be the only way to stop them.

At the Confidential Computing Summit this week in San Francisco, leaders from Google, Apple, Microsoft, AMD, Intel, Anthropic, and other organizations came together to discuss how to create a more secure, cross-platform system for AI agents that's safer than anything offered by a single vendor. 

The event's biggest takeaway? It's time to secure AI agents down to the chip level.

"Agents have to have cryptographically enforced identities," Aaron Fulkerson, CEO of Opaque and emcee of the event, told The Deep View. 

The biggest surprise of the event? Apple's Ivan Krstić, vice president of security engineering and architecture, showed up to talk about the next stage of Apple's Private Cloud Compute and how the company will use it to power the new Siri unveiled at WWDC 2026 earlier this month. 

"Apple doesn't speak at outside conferences," said Fulkerson, "so the significance of what they just announced with Siri sends a very clear message to the entire market that these systems are so intrinsically risky that you have to have a confidential, verifiable system."

Using a combination of open-source solutions and industry standards overseen by The Linux Foundation and the Confidential Computing Consortium, the two pillars that emerged from the event were:

• Hardware-enforced trust: The first step is to give every agent a unique identifier that can be traced to the chip level to show what they did, what they're allowed to do, and which data they touched. This will make agent behavior observable, trackable, and auditable at a level that meets the most stringent requirements of industries and enterprises, and prepare agents for compliance and regulation. 

• Agent verifiability: At the event, The Linux Foundation announced the Agent Name Service (ANS), which aims to provide every agent with something akin to a serial number, providing a tamper-proof record for tracking and registering agents. This will function similarly to the way the DNS system operates for servers on the internet. 

"There's multiple ways to do verifiable systems where you're creating a cryptographic signature to enforce and then prove that it was actually enforced," Fulkerson said. "The question becomes who does the signing? So that's why hardware makes sense. We can trust the hardware supply chain to do the signing, because of encryption keys."

It's no secret that AI agents have repeatedly shown they have the will and skills to bypass software-based rules and guidelines in order to accomplish a goal. That's their greatest strength: they will find a way to do the task that a human asked them to complete. It's also what makes them incredibly risky, and it's why AI pioneers such as Yoshua Bengio have been raising red flags about the imminent risk of these AI agents going rogue and doing things we never intended them to do. Insert your favorite sci-fi nightmare narrative, and it's easy to get a little uneasy. Much respect to the Confidential Computing Summit for bringing together industry heavyweights, academics, and people from various regulated industries for a meaningful dialogue focused on actionable next steps to make agents safe.

LINKS

• Meta poaches three founders from security startup Virtue AI

• US proposes partnership with EU on AI chip supply chains 

• AI revenue has reached a tipping point that justifies spend

• Scaled Cognition raises $100 million at $750 million valuation

• Apple hikes prices on devices amid AI memory chip crunch

• Meta eyes AI to review harmful content and ads

• LFM2.5-230M: Liquid AI's smallest model yet, built to run fast and on any hardware. 

• Synthesia Interactive Avatars: Generate a natural, real-time talking agent that can interact with users with conversation latency. 

• Zaro: A singular workplace to connect tools files and apps that live with your data. 

• Google Finance: Android users now have access to Google's finance app, with real-time market data and AI-powered “Key Moments.”

• Nvidia: Senior Scientist, Synthetic Data and Privacy

• Capital Group: AI AppSec Engineer Lead

• Luma: Research Scientist/Engineer – Foundation Model: Core Research

• AWS: Principal GenAI Specialist SA

GAMES

A QUICK POLL BEFORE YOU GO