SaaS Release — 2026-06-22 to 2026-07-05
Window: 2026-06-22 → 2026-07-05
Highlights
Section titled “Highlights”Invoke Kindo agents from any OpenAI Responses client
Section titled “Invoke Kindo agents from any OpenAI Responses client”The Agents API extension of the Responses API is complete. Any stored Kindo agent can be invoked from a stock OpenAI Responses client by passing its ID as the model (agent/<agent-id>) — the agent supplies the model, system prompt, and tools; the caller supplies the task. The capability set includes:
- Agent discovery alongside the model catalog (
include_agents=trueon the models list, with pagination for large agent fleets). - Live streaming of run activity — assistant text, reasoning summaries, and tool calls stream as standard Responses output items while the run executes.
- Polling for long-horizon runs, with run output populated incrementally as the run progresses.
- Full run transcripts readable through the Conversations API, including tool-call detail.
See the agent invocation guide to get started.
Integrations
Section titled “Integrations”No integration changes to report in this window yet.
Improvements
Section titled “Improvements”Resumable response streams
Section titled “Resumable response streams”If a streamed Responses API connection drops mid-response, clients can reconnect to the same response and replay the events they missed instead of starting over — every streamed event carries a sequence number, and browser EventSource clients resume automatically. Agent-run responses replay from the beginning of the run and then continue live. See Responses API streaming.
Refreshed API reference documentation
Section titled “Refreshed API reference documentation”The public API reference was audited against the live API surface at the conclusion of the Agents API project: streaming event vocabulary, conversation item types, message content parts (including file and image inputs), and run trigger semantics are all documented to match current behavior.
Anthropic endpoint compatibility with OpenAI-style clients
Section titled “Anthropic endpoint compatibility with OpenAI-style clients”The Anthropic-compatible messages endpoint accepts payloads produced by OpenAI-style gateways and SDK shims: system-role entries inside the messages array are honored as system prompts, and messages without an explicit role are treated as user messages. Clients that previously saw validation rejections now work unmodified.