Skip to content
Back to articles
devtools3 min

OpenAI Agents API: the Codex harness available for building your own agents

KGA IT
2026-10-123 min
AIOpenAIAgents APICodexAI AgentsGitHubDevelopers
Capa: OpenAI Agents API em beta pública

The announcement

On September 10, 2026, OpenAI launched the Agents API in public beta, giving developers the same "harness" (the layer that coordinates model calls, tools and context) and infrastructure that power Codex. The idea: instead of building an agent loop from scratch, you use the harness OpenAI maintains and focus on the tools, knowledge and workflows that make your agent useful.

How it works

  • OpenAI hosts and maintains the harness. It is based on the open-source Codex harness, so you can inspect the logic in the public code.
  • You choose where the agent runs: in an OpenAI-managed sandbox, on your own infrastructure or with sandbox partners.
  • OpenAI hosted sandbox: the same sandboxing infrastructure behind Codex and ChatGPT. It can be configured with your files, packages, skills and plugins so the agent can run code, work with files and produce artifacts.

Pricing

According to OpenAI, there are no additional fees for the Agents API: you pay only for the tokens and tools your agents use, per the API pricing page.

The repository behind it

The harness comes from openai/codex (Apache-2.0, about 128k GitHub stars). To try Codex locally before moving to the API:

npm install -g @openai/codex
# or
brew install --cask codex

For the Agents API itself, start with the overview and quickstart in OpenAI's docs, linked from the announcement.

Practical uses

  • Support or operations agents that need to run code, read spreadsheets or generate reports in an isolated environment.
  • Development task automation (reviews, migrations, tests) integrated with your systems.
  • Small teams that don't want to maintain their own sandbox infrastructure.

Caveats

  • Public beta means the API may change. Avoid depending on it for critical flows without a fallback.
  • Token costs can grow quickly with long-running agents. Set limits and monitor.
  • Security: even in a sandbox, review which data and credentials the agent can reach. Recent lab incidents of agents escaping sandboxes (OpenAI itself reported cases in July and September) show network isolation and monitoring matter.

Sources


Want to assess whether AI agents make sense for your processes? KGA IT offers a free consultation: kga-it.com.

Let's solve your technical challenges together.

KGA IT Solutions delivers AI, cloud, and DevOps expertise to address your specific challenges.

Contact Us