Public transcript

OpenAI President Greg Brockman: Our Plan To Merge Chat And Agents

By: Alex Kantrowitz
Length: 45 min
Language: English
Text from: the video's captions
Transcribed: 13 October 2026

Read the full transcript Open on YouTube

Notes in English.

Summary

OpenAI President Greg Brockman sits down with Alex Kantrowitz to discuss OpenAI's plan to merge chat and agents into a single interface, bringing Codex-style agentic capabilities into ChatGPT. Brockman explains that the long-term trajectory is toward an interface that

Key points

Greg Brockman says OpenAI's long-term trajectory is simplification and unification, merging chat and agents so the interface melts away.
Codex is a general-purpose agent harness, not just for code, and its usage inside OpenAI now rivals Slack in penetration.
The 2023 plugins failed because models weren't ready, with only about three connectors and 2K to 4K token context; today's models handle hundreds of tools.
Trust is the key product feature of the agentic era and must be earned through oversight, supervision and operator control.
Brockman says there is still no wall in sight for model scaling, with landmark results following a smooth compute-driven path since the 1940s.
Compute will be the scarce resource, and every provider sells out all their compute as demand grows.
Prices for a fixed level of intelligence always come down over a year-long horizon, while frontier intelligence stays pricey.
About 230 million people per week use ChatGPT for health queries, and Brockman says AI-assisted health management will become standard.

Questions it answers

00:00Merging chat and agents

Is OpenAI building a super app where everything starts with a prompt in ChatGPT?

Brockman agrees that's a good perspective. The goal is an AGI-like assistant that takes goals and directions from you and accomplishes them, with the interface eventually melting away into a persistent entity you simply talk to.

The long-term trajectory is simplification and unification of the product.
Brockman wants an AI that constantly thinks about what it can do for you.

02:42From suggestions to actions

How does ChatGPT go from suggesting actions to actually doing them?

Brockman confirms the shift from conversation to intent and action, citing Codex as the existing proof. He gives the example of a colleague whose agent gathered dietary preferences from attendees and built a seating chart for an event.

Codex is described as a general purpose tool-using harness, not just about code.
Trust is earned through tools, control, oversight and supervision.

04:41Why plugins failed

Why didn't the early attempts at tool use work?

The 2023 plugins failed because the models weren't ready: only about three connectors could be exposed at a time, with roughly 2K to 4K token context. Now models have huge context windows and can access hundreds of tools and whole file systems.

Models can now access hundreds of tools instead of three.
Brockman compares early models to 1960s computers with tiny memory.

06:33Trust and delegation

How much control will the agent have, like sending emails on your behalf?

It depends on setup and trust: the agent may draft emails and ask permission to send, or in a higher-trust world it drafts and sends them itself. Brockman stresses that trust is earned through oversight, supervision and control given to the operator.

Trust is a key product feature and differentiator.
Delegation levels depend on how much trust you've built with the system.

07:50Ecosystem and apps

Will ChatGPT face resistance from apps it needs to act through?

Brockman says it's not theoretical: Codex usage is exploding, with penetration inside OpenAI comparable to Slack. Partner companies responded positively to OpenAI training its AI to use their software, because blocking AI means declining rather than thriving.

Codex usage within OpenAI now rivals Slack.
Partners see AI training on their software as an opportunity, not a threat.

11:00Operating system or assistant

Is OpenAI trying to become the operating system for all apps?

Brockman says you could describe it that way, but he thinks of it as the ideal interface to a personal AGI: an assistant you talk to, which has its own computer plus delegated access to your systems, similar to how you already work with human co-workers.

Brockman prefers thinking of a personal AGI you talk to.
The ideal co-worker needs delegated access and context, like a human assistant.

13:10Competition with Siri and devices

How does OpenAI think about competing with Apple's new Siri, and its own devices?

Brockman avoids framing it as a direct fight, saying new capability levels let you rethink interfaces entirely. He acknowledges OpenAI's device work has been publicly reported, and Sam Altman told Kantrowitz in December that multiple devices are happening.

New capability means an opportunity to rethink everything.
Brockman won't confirm details but says multiple devices have been publicly reported.

18:15Natural voice interaction

What's the goal for voice models and natural conversation?

Current voice models chain speech-to-text, text-to-text and text-to-speech, with unnatural turn-taking hacks. The goal is a model that processes input and output simultaneously, like human conversation, allowing interruption and fluid real-time feedback.

Turn-taking hacks are unnatural and will be replaced.
Voice operation of Codex has been one of the most magical experiences.

Notes

Vision: Chat and agents merge

Greg Brockman frames OpenAI's goal as building toward AGI, with conversational intelligence being only one part of what people need to get work done.
The ideal interface is almost no interface: a persistent entity you talk to that accomplishes goals on your behalf.
The long-term product trajectory is simplification and unification, with fewer buttons, toggles and modes.

Codex as the agent proof point

Codex has code in the name but is a general-purpose tool-using harness; Brockman hooks his up to Slack, Gmail and calendar.
Example: a colleague on the comms team had Codex collect event attendees' dietary preferences and build a seating chart.
Codex usage inside OpenAI now has penetration comparable to Slack, and non-software work on it is growing exponentially.

Why plugins failed in 2023

Models could only handle about three connectors at a time with roughly 2K to 4K token context.
Today models can access hundreds of tools, whole file systems, and around 52 million tokens of context ("depends how you squint on it").

Trust in the agentic era

Trust is earned, not granted, through tools, control, oversight and supervision for the operator.
Depending on setup, an agent may draft emails and ask permission, or draft and send them in a high-trust configuration.

Voice and natural conversation

Current voice models chain speech-to-text, text-to-text and text-to-speech with unnatural turn-taking hacks.
The goal is a model that processes input and output simultaneously, allowing interruption like human conversation.

Scaling and compute

Brockman says there is still no wall in sight; landmark results since the 1940s neural nets and the 1959 perceptron follow a smooth compute-driven path.
Compute will be the scarce resource; every provider sells out their compute.
OpenAI has invested in its own chip program for multiple years, with more to announce soon.
Prices for a fixed level of intelligence always fall over a year; frontier intelligence stays pricey.

Differentiation and competition

Intelligence is not one-dimensional; domain depth (like AlphaGo's move 37 inspiring more Go players) leaves room for differentiation.
On Satya Nadella calling models a commodity, Brockman argues no layer of the stack loses value, citing rising prices even for previous-gen H100 chips.
On GPT-5.6 rumors: "You should always expect better, faster, smarter."

AI and health

About 230 million people per week use ChatGPT for health queries.
Brockman says his wife has health conditions he doesn't know how they'd manage without ChatGPT.
He calls applying AI to medicine a personal motivation and says stories like AI-assisted diagnoses will become standard.

Text from the video's own captions. Summary and notes written by AI from the transcript, so check anything important against the recording.