OpenAI launched the Agents API in public beta on 10 September 2026, giving all developers direct API access to the same managed orchestration infrastructure that powers its Codex coding agent — session management, context compaction, multi-agent coordination, and tool use included. The API carries no additional fees beyond standard token and tool costs, and the underlying harness is open-source at github.com/openai/codex.
Before this release, teams building long-running business agents had to write and maintain their own orchestration, context management, and failure-recovery logic. The Agents API eliminates that infrastructure build by providing it as a managed service, with agent sessions running in OpenAI-managed sandboxes or in partner environments.
What the Agents API Provides
The Agents API gives developers three infrastructure capabilities that previously required custom engineering. First, automatic context compaction: when an agent session approaches the context window limit, the API compacts earlier context automatically without developer intervention. Second, multi-agent coordination: the max_concurrent_subagents parameter is configurable per session, allowing a single API call to orchestrate parallel subagents on separate subtasks. Third, tool integration: the API supports MCP (Model Context Protocol), custom functions, and built-in tools including web search.
Compute runs in OpenAI-managed sandboxes or in 9 partner environments: Blaxel, Cloudflare, Daytona, DigitalOcean, E2B, Modal, Oracle, Runloop, and Vercel. The default model in the API’s reference example is gpt-6-astra, though the API supports multiple models.
Customer Results: 60–86% Improvements Reported
Three customers reported quantified results from the Agents API. Ciridae, an AI evaluation platform, reported that its evaluation score rose from 0.71 to 0.85 after migrating to the Agents API. “The subagent support in the API is great and drastically sped up our workflow. Previously it was pretty cumbersome to observe and orchestrate subagents in our old setup but the new APIs gave us a 4x latency reduction,” said Jack Weissenberger, CTO at Ciridae.
SafetyKit, a trust and safety platform, reported a 60% reduction in cost per case after moving its case review workflow to the Agents API, alongside lower latency and improved token efficiency, according to Bhavyansh Sabharwal, Member of Technical Staff at SafetyKit.
Hypha, a financial services AI provider, reported an 86% reduction in failed agent responses. “By separating the agent harness from the sandbox, we reduced failed agent responses by 86%,” said Serhii Shchoholiev, Lead Engineer at Hypha. Nash.ai, a logistics platform, uses the Agents API to run “thousands of long-running AI agents managing hundreds of millions of deliveries,” according to OpenAI’s announcement.
Why This Changes the Calculus for Business Software Buyers
The practical significance for businesses choosing software is supply-side: the Agents API lowers the cost and engineering time for software vendors to add genuine agentic capabilities to their products. Before September 10, embedding a reliable long-running agent into a CRM, project management tool, or HR platform required a custom infrastructure build worth months of engineering time. The Agents API delivers that infrastructure at token-cost prices.
The API also supports MCP — the open standard for connecting agents to external tools, backed by both OpenAI and Anthropic — a signal that the MCP ecosystem is now a realistic integration layer for the broader enterprise software market. For business buyers evaluating AI tools, agentic capabilities that were six to twelve months away on vendor roadmaps are now measurably faster to ship. The best AI agents for business in 2027 will increasingly run on this class of managed infrastructure rather than vendor-built orchestration layers.
For Context: OpenAI’s interest in agent infrastructure goes beyond the Agents API. In July 2026, researchers found that OpenAI agents used a German-language wiki to communicate with each other — a research finding on emergent multi-agent behaviour, distinct from today’s commercial product launch. AI agent adoption is already measurable across engineering teams: 8 in 10 engineers now use AI agents daily, according to Temporal’s 2026 developer report.
Our Take: OpenAI just handed software vendors a Codex-grade agent factory at token-cost prices. The cost and reliability numbers from Ciridae, SafetyKit, and Hypha are credible and large — 60–86% improvements in first-generation production deployments. The real beneficiaries are SaaS vendors who can now ship genuine multi-step agentic features without a seven-figure infrastructure build. Expect 2027 B2B software pricing and feature lists to reflect it.
