Skip to main content
AGENT cloud agents are Aria Compute’s cloud agent product. They serve the AGENT API in the OpenAI beta Agents shape: you create agents, manage sessions, and persist state, with memory in Postgres + pgvector. Turns run under DeepSeek Harness (dsh) by default, with native available for self-hosted deployments. Call the cloud entrypoints directly (https://agent.ariacompute.com); all five SDKs are cloud clients and embed no runtime or provider key. SDK memory backends are cloud, local (aria memo, SQLite) or both.

Agent concepts

The AGENT API exposes the OpenAI Agents resource shape. The core entities: The resource shape also includes events (including the agent.turn.approval.* and agent.turn.usage frames), projected session titles (child sessions readable through their parent) and subagents for multi-agent delegation. Streams terminate on agent.turn.completed or agent.turn.failed; there is no [DONE] sentinel.

Cloud entrypoints

The cloud AGENT API base URLs are:
  • International: https://agent.ariacompute.com
  • China: https://agent.ariacompute.cn
The management plane covers health, API keys and memory. The data plane authenticates with Authorization: Bearer <key> (or ApiKey <key>) plus OpenAI-Beta: agents=v1; endpoints include POST /v1/agents, POST /v1/agents/sessions, and POST /v1/agents/sessions/{session_id}/events[/stream]. Read-only governance is available at GET /v1/governance/status, /usage and /audit.

SDK access

The AGENT API is also callable through the JS and Python SDKs, which mirror the OpenAI Agents SDK. Define agents, bind tools, and start sessions and turns without hand-writing HTTP requests.

Memory, orchestration and approvals

The AGENT API adds memory, orchestration and approvals to the OpenAI Agents shape:
  • Memory: per-principal context_fragments in pgvector for cross-session retrieval; aria memo (SQLite) on device.
  • Subagents: delegate tasks within a turn; child sessions are isolated under <parent>::sub::<id>.
  • Teams and scheduled jobs: teams are validated task DAGs with dependency-gated claims; jobs fire at every_secs / at_ms and advance the cursor before execution — see Teams and Scheduled Jobs.
  • Approvals and permissions: one tool pipeline checks read_only / workspace_write / danger_full_access presets; with no approval provider the call is refused and the reason is fed back to the model, never silently allowed.
  • Tool families and browser: fs / shell / jobs / terminal / skills / web / ask / plan / todo / memo share one pipeline; the browser sandbox offers a six-image catalog.
  • Governed execution: egress policy, audit events, metering and credential substitution form four fail-closed seams; proxy enforcement is the default, with an optional network mode that requires a sidecar.

How to run

  • Cloud: call the entrypoints listed above; all five SDKs are cloud clients.
  • AGENT Runtime: self-hosting is optional (aria-agent serve --bind 127.0.0.1:8848). Configure the model with GATEWAY_BASE_URL / GATEWAY_API_KEY / GATEWAY_MODEL (no OPENAI_* fallback) and AGENT_HARNESS=dsh by default. Keep clear of aria-engine (8080 / 8010 / 8011) and aria-router (8899) ports.
  • Related: AGENT API introduction, ROUTER decision routing (the agent layer inside the gateway), AFM foundation models.

See also