Databricks GPT-5.5: What Enterprise Agent Workflows Need Operationally

Key Takeaways

• •OpenAI announced on May 15, 2026 that Databricks is making GPT-5.5 available for customer agent workflows.
• •The headline numbers are useful but not sufficient: OpenAI says GPT-5.5 reached 50% accuracy on Databricks' OfficeQA Pro benchmark and reduced errors by 46% compared with GPT-5.4.
• •The practical story is governance. Databricks positions Unity AI Gateway, AgentBricks, and the Agent Supervisor API as the control layer for model calls, tools, permissions, traces, and review.
• •Teams should verify actual GPT-5.5 availability, model ID, region support, and contract terms inside their Databricks workspace before planning production use.

Practical Analysis

This update matters because enterprise agent workflows usually fail in the unglamorous parts: scanned PDFs, old files, long documents, tool permissions, and audit trails. GPT-5.5 appears to improve the parsing-heavy and retrieval-heavy side of these tasks, but the operating model still decides whether the result can be trusted.

Databricks' Supervisor API lets a team define a model, tools, and instructions in a single request, while Databricks manages the loop of model calls, tool selection, tool execution, and final response synthesis. Unity AI Gateway is the governance layer for agents, LLM endpoints, MCP servers, and coding agents. In practice, the strongest pilot is not a broad autonomous agent. It is a bounded workflow with source evidence, tool traces, clear user permissions, and human review.

The main caveat is availability. OpenAI's customer story says GPT-5.5 is being made available through Databricks workflows, but the Databricks Supervisor API documentation checked for this draft still listed supported model parameters up to GPT-5.4. Treat the announcement as a product direction and validate the actual workspace model catalog before implementation.

Checklist

• □Is Unity AI Gateway enabled in the relevant Databricks account or workspace?
• □Is Supervisor API or AgentBricks available in the target region and Preview configuration?
• □Does the model catalog show the exact GPT-5.5 model string your app will call?
• □Are Unity Catalog objects, functions, Genie Spaces, Knowledge Assistants, Apps, and MCP connections permissioned separately?
• □Are traces written to Unity Catalog tables with `trace_destination`?
• □Do external connector and MCP calls require explicit approval where needed?
• □Is success measured by error rate, review time, rework, latency, and cost rather than generated volume?

Sources

• •OpenAI, Databricks brings GPT-5.5 to enterprise agent workflows: https://openai.com/index/databricks/
• •Databricks Docs, Supervisor API (Beta): https://docs.databricks.com/aws/en/generative-ai/agent-bricks/supervisor-api
• •Databricks Docs, Unity AI Gateway: https://docs.databricks.com/gcp/en/ai-gateway
• •Databricks Docs, Use Supervisor Agent to create a coordinated multi-agent system: https://docs.databricks.com/aws/en/generative-ai/agent-bricks/multi-agent-supervisor
• •Databricks GitHub, databricks-ai-bridge: https://github.com/databricks/databricks-ai-bridge
• •Databricks GitHub, Databricks CLI: https://github.com/databricks/cli