PulseFlow Tecnologia
← Insights
·6 min read·market-analysis·openai·codex·agentkit
PulseFlow Tecnologia

Is OpenAI turning into a software execution platform?

In a little over a year, OpenAI shipped four pieces that, looked at separately, seem like distinct products - Codex, the Responses API with Agents SDK, AgentKit, and the internally documented practice of harness engineering. Looked at together, they add up to something closer to a software execution platform than to a model provider that also sells a few tools.

Codex: an agent that already runs isolated, with no internet access

Codex is built on codex-1, a version of o3 optimized specifically for software engineering, trained with reinforcement learning on real-world coding tasks, to produce code that "closely mirrors human style and PR preferences" and runs tests iteratively until it passes. The most revealing detail is operational: the Codex agent runs entirely inside an isolated cloud container, with internet access disabled during execution - it only sees the repository's code and the dependencies pre-installed via a setup script. By OpenAI's own account, the most common internal use case is exactly what you'd expect from an engineering tool, not a chatbot: OpenAI's own engineers use Codex to offload repetitive, well-scoped tasks - refactoring, renaming, writing tests.

Responses API and Agents SDK: infrastructure, not just a model

The Responses API combines the simplicity of Chat Completions with the tool-use capabilities of the Assistants API, letting developers solve complex tasks with multiple tools and model turns in a single call - including web search, file search, and computer use natively. The Agents SDK, open source and provider-agnostic (works with 100+ other LLMs beyond OpenAI's own), formalizes concepts like handoffs between agents - transferring control from one agent to another within the same workflow. That's not a productivity tool: it's multi-agent orchestration infrastructure, the kind of piece that normally comes from a platform vendor, not a model vendor.

AgentKit: closing the loop from idea to production

AgentKit directly attacks the problem OpenAI itself describes: "building agents meant juggling fragmented tools - complex orchestration with no versioning, custom connectors, manual eval pipelines, prompt tuning, and weeks of frontend work before launch." The answer is an integrated set: Agent Builder (a visual canvas for creating and versioning multi-agent workflows), Connector Registry (a central place for admins to manage how data and tools connect across OpenAI products), and ChatKit (a toolkit for embedding customizable chat-based agent experiences into any product).

What this means

None of these four pieces alone proves OpenAI has become a "software execution platform." But the combination is hard to explain any other way: a coding agent that runs isolated and tests its own work (Codex), an API that orchestrates multiple tools and agents in a single flow (Responses API + Agents SDK), and a product that versions and connects entire workflows end to end (AgentKit) together describe the full cycle of building, running, and operating software - not just generating text or code from a prompt. OpenAI still sells model access, but more and more of the delivered value sits in the execution layer around it.

Sources