DevDay 2025: AgentKit and the Platform Play
OpenAI's DevDay on 6 October 2025 introduced AgentKit — a visual Agent Builder, the embeddable ChatKit UI, extended Evals and a Connector Registry — plus apps that run inside ChatGPT via an MCP-based Apps SDK. We summarize what shipped, what is beta, where lock-in sits, and how agent teams should frame the build-vs-buy decision.
Agent tooling is fragmented
Building a production agent in October 2025 means assembling a stack from parts: an orchestration framework for control flow, a hand-built chat frontend, a separate eval harness, custom connectors for every internal system. Each part carries its own versioning, its own failure modes, its own maintenance cost. Integration work dominates; the model call is the smallest piece.
On 6 October 2025, OpenAI addressed this fragmentation at its third DevDay in San Francisco. The keynote numbers frame the ambition: 4 million developers build on OpenAI, ChatGPT reaches 800 million weekly users, and the API processes over 6 billion tokens per minute — up from 300 million per minute in 2023. Against this backdrop the company announced AgentKit, a toolchain for agents, and apps that run directly inside ChatGPT.
What DevDay 2025 announced
AgentKit is the umbrella term for four building blocks. Agent Builder is a visual canvas for composing and versioning multi-agent workflows; it launched in beta. ChatKit is an embeddable, brandable chat UI component; generally available. Evals gained four new capabilities; generally available. Connector Registry is a central admin panel for data sources and MCP servers; beta rollout for API, ChatGPT Enterprise and Edu customers with the Global Admin Console. All four are included in standard API model pricing. DevDay also brought the Apps SDK preview, Codex general availability, GPT-5 Pro in the API and gpt-realtime-mini.
AgentKit at a glance
Agent Builder is the visible centerpiece. Workflows are node graphs: agents, tools, guardrails, conditional branches. The canvas supports preview runs, inline eval configuration and full versioning, and workflows can be exported as Agents SDK code in Python or TypeScript. Guardrails ship as an open-source layer against PII leakage and jailbreaks. OpenAI quotes Ramp with 70 percent shorter iteration cycles and an agent live in two sprints instead of two quarters. These are vendor-selected numbers, not independent benchmarks; an eight-minute live demo is not a production deployment.
| Component | Status on 6 Oct 2025 | Purpose |
|---|---|---|
| Agent Builder | Beta | Visual canvas: drag-and-drop nodes, guardrails, preview runs, versioning |
| ChatKit | Generally available | Embeddable chat UI: streaming, threads, thinking states |
| Evals (new capabilities) | Generally available | Datasets, trace grading, prompt optimization, third-party model support |
| Connector Registry | Beta rollout | Central governance of connectors and MCP servers across ChatGPT and API |
Evals move into the workflow
The Evals additions are the least spectacular and the most substantial part. Datasets combine automated graders with human annotations. Trace grading assesses complete agent runs end to end and locates the failing step. Automated prompt optimization feeds grader output back into prompts. Third-party model support lets the platform evaluate models from other providers.
What this does not do: evals do not replace domain expertise. Trace grading shows where a workflow fails, not what correct means for your process. Someone still has to define graders, label edge cases and maintain datasets. That work does not disappear — it moves from custom scripts into a hosted UI with versioning.
Apps in ChatGPT and the Apps SDK
The second announcement is about distribution. Apps run inside ChatGPT conversations with interactive UI — maps, playlists, design previews. The Apps SDK, released in preview, builds on the Model Context Protocol (MCP), the open standard Anthropic introduced in November 2024, and extends it with an interface layer. Launch partners: Booking.com, Canva, Coursera, Expedia, Figma, Spotify and Zillow. Notable for European teams: at launch, apps are not available in the EEA, Switzerland or the UK.
App submissions and a directory are announced for later in 2025; monetization guidance is pending, with the Agentic Commerce Protocol as the intended payment path. The offer to developers is reach — 800 million weekly users. The ranking and review criteria belong to OpenAI.
The platform play
The pattern is a classic platform play. AgentKit competes with orchestration frameworks, chat-UI libraries and standalone eval tooling — categories where OpenAI's own API customers operate. Apps in ChatGPT compete with the browser as a distribution channel. The GPT Store attempted this in January 2024 as a separate storefront; apps now render third-party UI directly in the conversation.
Two design decisions work against lock-in: Agent Builder exports workflows as Agents SDK code, and the Apps SDK is open source on MCP, so apps can run on any platform that adopts the standard. Both mitigate coupling; neither removes it. The hosted canvas, ChatKit's backend and the Connector Registry live in OpenAI's platform, and workflow execution runs on OpenAI models via the Responses API.
Build or buy for agent teams
Our build-vs-buy reading at Blue IT Systems reduces to three questions. First: where does workflow state live — in your repository or in a vendor console? Visual canvases accelerate prototypes; code remains the unit of review, testing and audit. Second: is the exit path real? Export the workflow and run it without the platform before you depend on it. Third: which data may pass through hosted connectors at all? For German enterprises that is a GDPR and data-processing-agreement question before it is a feature question.
One structural limitation remains: third-party model support in Evals covers grading, not execution. Agent Builder workflows execute on OpenAI models. Teams with a multi-model strategy — routing, fallback, cost tiers — still need a neutral orchestration layer. AgentKit does not provide one, and is not designed to.
Outlook from October 2025
Our expectations, written in October 2025. Visual builders will win the prototyping phase; we doubt they will win production, where change management, code review and reproducible deployments favor exported code plus evals. The quiet winner of this DevDay is MCP: when the largest AI platform ships its app layer on a competitor's open standard, the standard has won. Competing canvases from other vendors are a matter of months.
Open questions: whether apps reach the EEA, how monetization is split, and whether hosted workflow builders survive contact with enterprise governance. Our working stance: use AgentKit as a fast prototyping lane with a mandatory export test, adopt the eval capabilities first, and keep orchestration portable. Whether these predictions age badly is something the evals will tell us.
Sources
- OpenAI – Introducing AgentKit (6 Oct 2025)
- OpenAI – Introducing apps in ChatGPT and the new Apps SDK (6 Oct 2025)
- TechCrunch – OpenAI launches AgentKit to help developers build and ship AI agents (6 Oct 2025)
- TechCrunch – Sam Altman says ChatGPT has hit 800M weekly active users (6 Oct 2025)
- The Verge – OpenAI will let developers build apps that work inside ChatGPT (6 Oct 2025)