Loading skills in the main agent loop

[NCCYUNSONG]

First of all, thank you for the excellent work on the OpenAI Agents SDK. It provides a very clean and flexible foundation for building agent-based systems.
Using skills through Codex or the hosted shell runtime introduces extra steps and is not convenient in many cases.
For many workflows (e.g., local development, healthcare support, research experiments, multi-agent prototyping), developers often need to use local scripts that they want to expose to agents.
Direct local skill support would simplify development and reduce friction.
For example:

from agents import Agent
agent = Agent(
    name="Shell Agent",
    model="gpt-5.2",
    instructions="Use available skills to solve tasks.",
    skills=[exp1, exp2]
)

Thank you again for building and maintaining the Agents SDK. Looking forward to future improvements and discussions around skill extensibility.

[seratch]

Thanks for sharing the idea. This SDK's main agent loop is not currently optimized to behave like a coding agent (or in other words, an agent loop that mainly works with shell runtime), so we don't have immediate plans to enhance the main agent loop in the direction. That said, we'd love to learn from any suggestions / implementation ideas.

[NCCYUNSONG]

Thanks for the clarification — that makes sense.

I would like to clarify that my use case is not related to building a coding agent or a shell-based development workflow. The motivation mainly comes from healthcare and research systems, where data security and compliance requirements often require local processing.
In these environments, developers often need local utilities or domain skills, for example:

clinical data processing scripts
medical guideline retrieval
survival analysis or statistical scripts
pathology or imaging preprocessing utilities

These skills are typically implemented as local Python scripts or modules and are frequently updated during development.
Being able to expose such local skills directly in the SDK would make it easier to integrate them into agent workflows without requiring the hosted shell runtime or uploading them as remote skills.

[seratch]

Thanks for sharing more details. If those utilities are implemented in Python and the module encapsulates the relevant knowledge and requirements, simply using them as function tools should work well: https://openai.github.io/openai-agents-python/tools/#function-tools

I understand that adopting the agent skills format could make things more portable across systems and projects, but if you do not specifically need markdown documents like SKILL.md, attaching your Python scripts or modules should already work with this SDK.

[NCCYUNSONG]

Thanks for the explanation and for pointing to the function tools documentation.
That makes sense- using function tools worked well for me in many cases. I think there will still be more possibilities to explore if skill support is added in the future.
I appreciate the clarification and look forward to future developments around skills in the SDK.

[gauravdhiman]

Thanks for opening this request and for the discussion. I think, in general, this would be a great capability if agents can be enabled to load skills. As @NCCYUNSONG mentioned, this may not only be for coding cases but for many other scenarios where skills would be really helpful. Skills are quickly becoming a portable way to provide domain knowledge and enable agents with new capabilities. Many other frameworks now support them, like LangChain, Agno and Google ADK. It would be great to have similar capabilities in the OpenAI Agents SDK.

I strongly feel the ecosystem of skills is becoming stronger day by day and is the most portable way to steer agents in a specific domain with domain specific knowledge and processes. @seratch if you plan to not support it, may I know the rationale behind it, especially when it's becoming a standard in the ecosystem? Having that insight will be really valuable. Hope to see OpenAI Agents SDK start supporting Skills natively.

[seratch]

Effectively and efficiently loading skills requires different model behavior. It consumes many turns for a single skill execution, so it does not fit naturally into this SDK's main agent loop at the moment. For this reason, skills are currently supported within the following tools:

• Shell tool: https://github.com/openai/openai-agents-python/blob/main/examples/tools/local_shell_skill.py
• Codex tool: https://github.com/openai/openai-agents-python/blob/main/examples/tools/codex.py

These tools can effectively use skills as part of their task execution.

Please note that we may provide additional ways to support skills in the future, so this is not where we stop. I just wanted to share what is available right now.