Databricks Puts GPT-5.5 to Work on Enterprise Agent Workflows

Databricks is making GPT-5.5 available for customer agent workflows after the model set a new state of the art on OfficeQA Pro, the company's benchmark for complex enterprise document tasks. In the agent-harness setting, GPT-5.5 reduced errors by 46% compared to GPT-5.4 and became the first model to surpass 50% accuracy on OfficeQA Pro.

The announcement matters because the benchmark targets exactly the tasks that break production agent systems in real enterprises: parsing, retrieval, and grounded reasoning across scanned PDFs, legacy files, and long-context documents. That is where Databricks' customers run their agents, and it is where earlier models have consistently failed.

A benchmark built for the messy reality of enterprise documents

OfficeQA Pro is not a synthetic test of general reasoning. It evaluates how models handle workflows involving large numbers of scanned or legacy enterprise documents. The difficulty is structural. Small extraction errors during parsing can cascade downstream through the rest of a workflow, corrupting everything an agent does afterward.

"Once you can't extract a certain digit or number, that changes the entire trajectory of what the agent works with," explains Arnav Singhvi, a Research Engineer at Databricks.

That single missed digit can invalidate a retrieval step, misdirect a multi-step plan, and produce a confident but wrong answer. This failure mode is why document-heavy enterprise deployments have remained hard even as general-purpose model capabilities have improved. A model that scores well on broad academic benchmarks can still fall apart when it meets a low-quality scan of a 2009 expense report.

Databricks saw the largest gains from GPT-5.5 precisely in these parsing-heavy workflows. "Earlier models like 5.4 were unable to parse all the digits correctly, but it seems like 5.5 has a step-function lift in parsing older documents and scanned PDFs," Singhvi says.

That description — a step-function lift rather than an incremental improvement — is the core claim behind the 46% error reduction. The model did not just get somewhat better at the same tasks. It crossed a threshold in a capability that had been a bottleneck for entire classes of enterprise automation.

Fewer detours, tighter orchestration

Parsing accuracy was not the only gain. The Databricks team also observed improvements in orchestration across multi-step tasks, the second half of what makes document agents fragile.

"One thing that we saw with 5.4 is sometimes it would go on these unnecessary search detours, and that would cause very inefficient trajectories," Singhvi says.

Those detours are costly in production. Every unnecessary search step adds latency, consumes compute, and increases the surface area for errors to compound. Compared to earlier models, GPT-5.5 was more reliable at retrieving relevant context and completing complex workflows without additional supervision. For enterprises, that last phrase carries weight: less supervision means fewer human checkpoints, less human-in-the-loop overhead, and agent workflows that can actually run end to end.

The combination matters. Accurate parsing without reliable orchestration produces correct fragments assembled into inefficient or broken workflows. Reliable orchestration without accurate parsing produces confident agents built on corrupted inputs. OfficeQA Pro tests both together, and GPT-5.5 is the first model to clear 50% accuracy on it.

"Codex with 5.5 is now state-of-the-art amongst all the agents and models out there," Singhvi said.

How customers get it

Databricks is not offering GPT-5.5 as a standalone endpoint for its own sake. The company makes the model available through AI Unity Gateway, where customers use it inside workflows built with AgentBricks and the Agent Supervisor API.

The architecture is specific. In these systems, GPT-5.5 orchestrates parsing, retrieval, and execution across specialized agents. It acts as the coordinating layer — the component that decides what to extract, what to retrieve, and which specialized agent handles which step — rather than as a monolithic assistant doing everything itself.

That division of labor reflects how Databricks sees enterprise agents evolving: specialized agents for well-defined tasks, with a strong general model in the supervisor role. The 46% error reduction and the first-above-50% result on OfficeQA Pro suggest the supervisor role is where model quality pays off most, because a supervisor's mistakes propagate through every agent beneath it.

"We're going to see a lot of customers using AgentBricks and Agent Supervisor API for custom agent workflows," Singhvi says. "Having GPT-5.5 supervise these workflows is really exciting."

Why the stakes are concrete

The market context is straightforward. Enterprises have spent the past two years piloting document agents, and the pilots routinely stall on exactly the failure modes OfficeQA Pro measures: scanned PDFs that parse badly, legacy files in outdated formats, long documents where retrieval misses the relevant passage. A benchmark result showing a 46% error reduction on those specific workflows, from a vendor that ships the surrounding agent infrastructure, is a signal that the production gap may be narrowing.

Singhvi's assessment of the model's impact on Databricks' own work is blunt: "GPT-5.5 has been great in terms of knowledge lift. It's a step size function change in terms of doing knowledge work for us."

What comes next follows directly from the deployment path. With GPT-5.5 now live through AI Unity Gateway for AgentBricks and Agent Supervisor API workflows, the measure of the model shifts from benchmark scores to whether the parsing and orchestration gains hold up in customer environments full of documents that no benchmark fully anticipates. The first sign to watch: whether workflows built on scanned and legacy documents, the hardest cases in OfficeQA Pro, run in production with materially less human supervision than they required under GPT-5.4.