OpenAI Responses API Long-Running Agent Support Update

OpenAI has significantly updated its Responses API, introducing three new features to address the challenges of long-running agents. With Server-side Compaction, Hosted Shell Containers, and the Skills API, agents can now run continuously for hours or even days.

Server-side Compaction Resolves “Context Forgetting”

Traditional AI agents faced issues with token limits during long task executions, leading to the loss of past inferences. OpenAI’s Server-side Compaction fundamentally resolves this problem.

This feature is not just a simple history deletion but summarizes the agent’s past actions in a compressed state, retaining important context while removing noise.

# Enable automatic compaction
curl -X POST https://api.openai.com/v1/responses \
  -H "Authorization: Bearer $OPENAI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-5.2",
    "store": false,
    "compaction": {
      "mode": "auto",
      "threshold": 100000
    }
  }'

According to the demonstration data from the e-commerce platform Triple Whale, agent Moby maintained accuracy while operating in a session with 5 million tokens and 150 tool calls. (Source: venturebeat.com)

Hosted Shell Containers Provide a Complete Execution Environment

The new Shell Tool offers an OpenAI-hosted Debian 12 environment with the container_auto option. This is not just a code interpreter but provides each agent with a dedicated terminal environment.

# Example usage of Hosted Shell Container
import openai

response = openai.responses.create(
    model="gpt-5.2",
    tools=[{
        "type": "shell",
        "shell": {
            "container": "container_auto"
        }
    }],
    messages=[{
        "role": "user", 
        "content": "Create a Python data transformation script and save it to /mnt/data"
    }]
)

This environment comes with the following pre-installed:

  • Python 3.11, Node.js 22, Java 17, Go 1.23, Ruby 3.1
  • Persistent storage via /mnt/data
  • Internet connection for library installation and external API integration

Data engineers can implement high-performance data processing tasks in these hosted containers without building ETL middleware. (Source: venturebeat.com)

Skills API for Standardized Agent Capabilities

The Skills API is a new standard developed by the OpenAI team, allowing agent capabilities to be defined in a structured format. Compared to traditional custom function definitions, this approach improves reusability and maintainability.

{
  "type": "skill",
  "skill": {
    "name": "data_analysis",
    "description": "Perform data analysis and visualization",
    "parameters": {
      "data_source": {"type": "string"},
      "analysis_type": {"type": "string", "enum": ["statistical", "trend", "correlation"]}
    }
  }
}

The Skills standard enables developers to share capabilities across multiple agents and reduce repetitive prompt work. (Source: openai.com)

Implementation Architecture Details

In OpenAI’s official announcement, the internal implementation of Server-side Compaction is described as “summarizing the agent’s past actions in a compressed state, retaining important context.” However, specific details about the compression algorithm and memory management are not disclosed.

According to Microsoft Azure documentation, compaction items are encrypted and not in a human-readable format, suggesting that OpenAI has developed its own state compression technology. (Source: learn.microsoft.com)

Summary

  • Using Server-side Compaction, long-running agent sessions with 5 million tokens can be executed while maintaining accuracy
  • Hosted Shell Containers provide persistent storage via /mnt/data, allowing for data retention between sessions according to company policies
  • The Skills API standard enables significant reductions in prompt development time and allows for capability sharing across multiple agents
  • With the container_auto option, high-performance data processing pipelines can be instantly built without ETL middleware construction