Agentic Execution Environment

7/20/2026

This post introduces custom repository rules for Codex Code Review, allowing developers to define project-specific guidance within `AGENTS.md`. This feature enhances the agentic execution environment by equipping coding agents with the ability to understand and enforce nuanced project invariants, compatibility requirements, and data boundaries during code reviews. This reduces the burden on human reviewers, accelerates the development cycle by automating context-aware checks, and enables agents to operate more effectively on complex, long-horizon coding tasks by internalizing team knowledge.

6/26/2026

This post introduces Secure MCP Tunnel and Harpoon, which significantly enhance the agentic execution environment by enabling secure, outbound-only connections from private enterprise networks to OpenAI products. This allows agents to access private data and services (via MCP or REST APIs) without requiring public endpoints or broad network access, thereby preserving security boundaries and simplifying integration for developers. It also emphasizes the inspectable, customer-controlled nature of the tunnel client and extends the connectivity model beyond MCP to general REST APIs.

6/23/2026

This post details the technical enhancements to the Agentic Execution Environment, specifically focusing on the 'Remote' functionality within the ChatGPT mobile app. It introduces the concept of the phone as a 'control plane' for remote development machines, rather than a terminal. Key technical additions include: remote host connections, worktrees for isolated development, goals for durable objectives, side chats for branching thought processes, inline code review with diff inspection and commenting, queued vs. steered prompts for controlling agent execution flow, attachment handling for context, skills and plugins integration, archived chats for organization, and granular security controls for command approvals. It also outlines workflows for release management, bug fixing, and general task management using these mobile-first engineering tools.

3/11/2026

This post highlights how developers are using the Responses API to build more sophisticated agentic systems. It details Raindrop AI's use of the API for monitoring and debugging AI agents in production, Repo Prompt's architecture for deep reasoning on complex data by separating context building from analysis, and Collxn's conversational interface for vinyl record collectors using tool calling. These examples showcase the evolution of agentic execution environments towards more reliable, complex, and interactive applications.

3/9/2026

This post details the integration of repo-local skills, AGENTS.md, and GitHub Actions to automate OSS maintenance for the OpenAI Agents SDK. It describes how skills encapsulate repository-specific workflows for tasks like code verification, documentation synchronization, release review, and PR drafting. The use of AGENTS.md enforces mandatory skill usage for critical steps, ensuring consistency and quality. Specific examples of skills for Python and TypeScript SDKs are provided, along with the structure of AGENTS.md for defining mandatory workflows and verification rules. The post highlights how this approach accelerates development throughput and improves the accuracy of recurring engineering work.

2/26/2026

This post introduces the integration of Codex with Figma via the Figma MCP server, enabling bidirectional UI generation. It details how to extract design context from Figma files (designs, makes, FigJam) to inform Codex for code generation, and conversely, how to render live UIs and convert them into editable Figma frames. This enhances the agentic execution environment by allowing agents to directly interact with and generate UI elements based on design specifications and to iterate on code-generated UIs within a design context.

2/23/2026

This post details an experiment demonstrating the capability of GPT-5.3-Codex to build a design tool from scratch over approximately 25 hours, using 13M tokens and generating 30k lines of code. It highlights the critical shift towards 'time horizon' in agentic coding, emphasizing the agent's improved ability to stay coherent, complete larger chunks of work end-to-end, and recover from errors. The post introduces the concept of 'durable project memory' using structured markdown files (spec, plan, implement, documentation) as a key technique for maintaining context and preventing drift in long-running tasks. It also showcases the agent loop (plan, edit code, run tools, observe, repair, update docs) and the practical application of Codex features like Skills, Automations, and Git worktrees in supporting these long-horizon tasks. The experiment underscores the move from micromanagement to delegation with guardrails for agentic work.

2/11/2026

This post introduces and elaborates on three new agentic primitives: Skills (reusable, versioned instructions), an upgraded Shell tool (hosted or local containerized execution environment), and server-side Compaction (automatic context window management for long runs). It provides practical tips and build patterns for effectively using these primitives together to enable long-horizon agents to perform real knowledge work, focusing on reliability, security, and developer experience.

2/4/2026

This post details 15 lessons learned from building ChatGPT Apps, focusing on the technical challenges and solutions for integrating UI widgets with the AI model in an agentic environment. Key contributions include strategies for managing context asymmetry between the UI, model, and user, emphasizing explicit information sharing and visibility. It introduces techniques like using `_meta` fields for UI-specific data, lazy-loading alternatives with aggressive front-loading, declarative UI context updates via `data-llm` attributes and Vite plugins, and mapping interaction paths. The post also covers UI design considerations for multiple display modes, UI consistency using the SDK UI Kit, language-first filtering via Lists of Values (LOVs), and leveraging files for richer interactions. Finally, it addresses production concerns such as Content Security Policies (CSPs) for secure iframe isolation and the use of app manifests for declaring allowed domains.

1/22/2026

This post introduces a systematic approach to testing and evaluating agent skills, particularly within the context of Codex. It details how to define success metrics before skill development, create skills with clear names and descriptions, manually trigger skills to uncover assumptions, and use small, targeted prompt sets for regression testing. The core contribution is the methodology for using Evals, which involves defining outcome, process, and style goals, and implementing lightweight deterministic graders and rubric-based scoring to ensure skill reliability and consistency.

1/11/2026

This post details how integrating OpenAI's Codex with JetBrains IDEs via the Model Context Protocol (MCP) dramatically enhances agentic capabilities. It allows agents to access IDE context such as file problems (compilation errors, warnings) and execute predefined run configurations (tests, linters, formatters). This integration provides agents with immediate, precise feedback loops, similar to human developers, leading to faster error detection, reduced iteration cycles, higher quality code suggestions, and better alignment with existing development workflows. The post emphasizes that context awareness, particularly through IDE integration, is crucial for agent effectiveness.

12/30/2025

This post details the 2025 evolution of OpenAI's agent capabilities, focusing on the shift to agent-native APIs and improved models. Key contributions include the Responses API for multi-modal inputs/outputs and reasoning controls, the Agents SDK and AgentKit for easier agent orchestration, and enhanced state management with Conversation State and Connectors. It also highlights the maturation of integrated tools (web search, file search, Code Interpreter, computer use) and platform features for scaling and reliability, such as prompt caching, background mode, and webhooks, all contributing to more robust production-grade agent development.

11/24/2025

This post introduces and defines ChatGPT Apps as a new way to integrate external products into ChatGPT conversations. It shifts the paradigm from porting entire products to exposing specific, well-defined capabilities (Know, Do, Show) that the model can orchestrate. The post provides practical guidance on selecting use cases, designing for conversational discovery, and defining clear APIs for these capabilities, emphasizing that the value lies in discrete operations that extend the model's reach into data, actions, and structured UIs, thereby enhancing agentic capabilities.

10/10/2025

This post highlights the pervasive use of Codex, an AI coding assistant, in building and executing various components for OpenAI DevDay 2025. It showcases how Codex was instrumental in controlling hardware (cameras, lights), developing rich user interfaces (beat pad, arcade games), rebuilding demo applications, generating diagrams, and managing documentation. Codex's capabilities in code generation, debugging, parallel task execution (via CLI and IDE extensions), and rapid iteration significantly accelerated development cycles and enabled the creation of complex features that would have been time-prohibitive otherwise. This demonstrates a significant advancement in the practical application of AI for agentic development and deployment.

9/22/2025

Introduces the `/v1/responses` API, designed as a stateful, multimodal, and efficient agentic loop. Key technical contributions include preserving the model's reasoning state across turns (improving benchmarks and cache utilization), emitting multiple output items (tool calls, structured outputs, intermediate steps) for better debugging and richer UIs, and integrating hosted tools (web search, image gen, MCP) for server-side execution. The API's polymorphic item structure clarifies the ordering of actions, and it's designed to be multimodal from the ground up, unlike previous text-centric APIs.