OpenAI Agents API: Managed Harness & Sandboxes Replace Custom Agent Frameworks

One Call, Fully Managed Framework

OpenAI has released a public beta of the Agents API, which essentially turns the harness that powers Codex into a managed cloud service. Developers specify a task, model, tools, and environment in a single API call, and the service runs a production-grade agent. The underlying logic for context management, tool scheduling, and sub-task coordination is fully handled by OpenAI.

const client = new OpenAI();

const session = await client.beta.agents.sessions.create({
  agent: {
    model: "gpt-6-astra",
    tools: [{ type: "mcp", server_label: "observability" }],
    multi_agent: { enabled: true, max_concurrent_subagents: 3 },
  },
  environment: { type: "openai_hosted" },
  input: "Investigate the 5xx spike in service-api over the last 30 minutes. Delegate sub-agents for deployment, error, and dependency analysis. Save findings and mitigation suggestions to /workspace/outputs."
});

Previously, building agents involved several pain points: context window limits, token-heavy tool calls, and custom sub-task coordination. The Agents API addresses these with:

Automatic context compression: when a session nears the context limit, the system automatically condenses earlier content, eliminating the need for custom compression logic for long-running tasks.

On-demand tool definition loading: tool search loads definitions only when needed, preserving model cache and reducing token consumption.

Programmatic tool calls supporting parallel execution and chained operations, returning only relevant results for large data volumes.

Multi-agent support: a primary agent decomposes tasks, while sub-agents run in parallel with independent contexts.

The model used is gpt-6-astra.

Sandbox Options

Different tasks require different compute, storage, and deployment models. The API offers three sandbox choices:

OpenAI-hosted sandbox, sharing the same infrastructure as Codex and ChatGPT.

Self-hosted infrastructure deployed in your own VPC.

Partner sandboxes from Blaxel, Cloudflare, Daytona, DigitalOcean, E2B, Modal, Oracle, Runloop, and Vercel.

The choice depends on requirements: OpenAI-hosted is simplest for rapid validation; self-hosted VPC suits production compliance and storage needs; partners vary in CPU, GPU, memory, cold-start latency, and cost trade-offs.

Early Customer Data

Several beta teams reported metrics:

Ciridae: evaluation score improved from 0.71 to 0.85, latency reduced 4x.

SafetyKit: case review cost dropped 60%, token efficiency significantly improved.

Hypha: separating harness from sandbox reduced agent failure responses by 86%.

Ciridae's CTO noted they previously spent extensive time optimizing sub-agent observation and orchestration; the new API delivers out-of-the-box gains. These improvements stem from no longer maintaining custom agent infrastructure, freeing focus for business logic.

Open Source, Not a Black Box

The Agents API is powered by the open-source Codex harness, with code publicly available on GitHub. OpenAI operates and maintains the managed service, while developers can inspect the core logic. The hosted service maintains transparency, avoiding vendor lock-in concerns.

Billing Reality

The release prompts industry jokes about "OpenAI killing 100+ startups," reflecting the platform-layer trend of turning infrastructure into services, forcing middleware companies to reposition. However, cost concerns exist: one user reported a single retry triggering 12 tool calls, with traces returned to context, multiplying the bill before the task completed. OpenAI states no extra fees during public beta—charges are based on token and tool usage only. Savings on framework maintenance may shift to runtime costs; developers must balance tool automation against cost control.

Advice for Developers

The cost to validate an agent idea is now unprecedentedly low. Teams without internal infrastructure can start directly on Codex's shoulders. Teams with existing custom frameworks should ask whether maintaining the harness is a core competency; if not, the managed service can save significant time.

The Agents API remains in public beta; trying it now is free aside from token and tool usage—just one API call away.

Large models continue to compress the application developer's space: first Models-as-a-Service (MaaS), then Harness-as-a-Service (HaaS), Bot-as-a-Service—pushing developers toward the application and user layers.