OpenAI 2025: Agents, Apps, Codex GA
OpenAI introduced at DevDay 2025 a comprehensive package for production-ready AI agents: AgentKit for building and deploying agents, Apps in ChatGPT including the Apps SDK, and the general availability of Codex for developer teams. These innovations raise questions about immediate usability, monetization, and EU availability.
Quelle: Custom Display
OpenAI AgentKit Overview
AgentKit is a toolbox that bundles the development, testing, and operation of agents. Core building blocks are the Visual Agent Builder, ChatKit, as well as evaluation features and Reinforcement Fine-Tuning (RFT) for robust quality in operation. The Agent Builder enables the design of workflows, versioning, and guardrails. ChatKit provides ready-made chat UI components for embedding in web/apps and is generally available. The Agent Builder starts in beta.
Apps in ChatGPT are standalone, interactive mini-apps that run directly in the chat. They are built with the new Apps SDK built, based on the open standard Model Context Protocol (MCP), and allows developers to define the app's logic and UI.
Codex is OpenAI's Coding Agent for software teams. It is now generally available and brings a Slack integration, its own SDK, and admin tools for enterprises.
Quelle: adasci.org
Architecture and functionality of the OpenAI Agents SDK, enabling the development of agentic AI applications.
On October 6, 2025, OpenAI announced Apps in ChatGPT and the Apps SDK. The first partner apps came from Booking.com, Canva, Coursera, Expedia, Figma, Spotify, and Zillow. Apps are 'today' available for logged-in ChatGPT users outside the EU on Free, Go, Plus and Pro. Further partners and an EU rollout are to follow. App submissions and monetization details are planned for 'later this year'.
The apps behave context-sensitively in chat, can be proposed when needed, and bring interactive elements such as maps, playlists, or presentation views directly into the chat. Media reports confirm the first integrations and describe the functionality similar to a 'mini-app' ecosystem ( The Verge, TechCrunch).
In parallel OpenAI introduced AgentKit: Agent Builder (visual orchestration), ChatKit (embedding), Connector Registry (data/tool governance), Evals extensions and RFT options. ChatKit and the new Evals are generally available. Agent Builder and the Connector Registry start in beta and can be activated via the Global Admin Console.
Codex went generally live – including Slack integration, Codex SDK (TypeScript first), new admin features, as well as customer examples and usage numbers. According to OpenAI, nearly all internal engineers use Codex; merged pull requests are rising 70% weekly, and GPT-5-Codex handled 40 trillion tokens within three weeks.
Strategic Analysis
Strategically, OpenAI opens two fronts: On one hand, ChatGPT becomes the platform where apps appear directly in the dialogue — a move toward 'context distribution,' because the assistant suggests suitable apps at the right moment. That lowers friction for users and offers developers reach to 'over 800 million ChatGPT users.' On the other hand, AgentKit addresses enterprise hurdles: governance, reproducible workflows, measurability, and faster UI embedding — in short, less tool fragmentation in the agent stack. Wired). Reuters emphasizes the focus on enterprise customers and partnerships ( Reuters). For developers, that means: less manual work on the agent frontend, clearer admin guardrails, and a potential distribution channel via the new app publication along with monetization and the announced Agentic Commerce Protocol.
Quelle: medium.com
Schematic representation of the interaction of agents and protocols, a core concept of agentic AI.
Quelle: YouTube
The keynote stream provides original statements on product strategy and demos.
Evidence: Apps in ChatGPT are live outside the EU on Free/Go/Plus/Pro; Apps SDK as an open MCP extension; App submissions and monetization 'later this year'; partner list and behavior of the apps. It is also evidenced that: AgentKit includes Agent Builder (Beta), ChatKit (GA), Connector Registry (Beta rollout), Evals extensions and RFT options; activation via the Global Admin Console possible. It is also evidenced that: Codex GA with Slack integration, Codex SDK (TypeScript), admin tools, as well as the aforementioned usage figures and customer examples.
Remaining unclear: concrete revenue share and fee structure for apps; OpenAI points to later details. The concrete EU timeline and which regulatory obligations must be met are also unclear. The claim 'This is only a rebranding of old plugins' is misleading. The new apps bring interactive UIs, are MCP-based, are contextually suggested in the chat, and receive an official publication and review process; that sets them apart substantively from earlier plugin experiments.
The Verge describes the Apps experience with concrete examples (Canva designs, Zillow maps) and classifies them as a broad app start in ChatGPT. TechCrunch and VentureBeat emphasize openness to third-party providers and the role of the Apps SDK, partly focusing on initial brand partners. Wired reflects the platform strategy ('ChatGPT as the operating system') – with opportunities but also dependencies on gatekeeping and policies. Reuters highlights the enterprise agenda and partnerships.
The fusion of ChatGPT and OpenAI technologies points to a future in which AI agents could control devices and applications.
Quelle: tomsguide.com
The fusion of ChatGPT and OpenAI technologies points to a future in which AI agents could control devices and applications.
Practical implications
When you build agents for customer deployment, AgentKit reduces time to live deployment: visual orchestration, standardized evals, guardrails, and an immediately usable chat UI via ChatKit. If you want to offer existing services in ChatGPT, the Apps SDK is the quickest path to distribution in chat — including app suggestions at the right time. If you want to automate developer workflows, Codex GA provides an integrated coding agent along with an SDK and admin controls.
Practical next steps: review documentation/examples, enable Developer Mode, and pilot a vertical use case as a Lean App/Agent; early consider privacy prompts, permissions, and future app policies. Plan for EU rollout scenarios with 'region fallback' and follow the announcements.
Quelle: YouTube
Open questions remain: What exactly do revenue share, fees, and the review process for apps look like, and what ranking criteria apply in the app directory? When does EU availability start concretely and with what granular data permissions per app? How quickly will the Agent Builder mature to GA, and when will the announced Workflows API be released? What guardrails will apply in the future for commerce in chat via the Agentic Commerce Protocol?
Conclusion: AgentKit cleans up the agent stack: less integration effort, more measurability, and faster embedding into real-world products. Together with Apps in ChatGPT, it creates a seamless path from idea to a distributed application within the chat. Those who pilot early, implement policies cleanly, and keep an eye on EU availability can gain an early lead in the design of agentic products.