Red Hat packages a ten-agent aircraft lease workflow for OpenShift AI

Red Hat has added an aircraft-lease processing quickstart to its OpenShift AI learning catalog. The deployable example takes PDF contracts through ten sequential agents, links extracted facts to source clauses and produces return, extension or buyout recommendations.

The useful part is not the aviation-specific interface. It is the packaged pattern for moving a document-heavy agent workflow from a diagram into an OpenShift deployment with an inference service, job queue, application state and an audit trail.

What the quickstart builds

The Helm chart deploys a Next.js frontend, FastAPI backend, background worker, PostgreSQL and Redis. LangGraph orchestrates the agents, while IBM Granite 3.3 2B Instruct runs through vLLM and LlamaStack on OpenShift AI's KServe layer.

The pipeline begins by classifying an uploaded contract, then extracts terms, maps obligations and reconciles operational data. Later stages calculate reserves, flag variances, assess return readiness and assemble an evidence pack before a decision-support agent produces its recommendation. A separate retrieval-augmented assistant answers questions about processed contracts with source citations.

Red Hat estimates 30–60 minutes for deployment, with a longer first run while the model downloads. The documented baseline is OpenShift 4.19 or later and OpenShift AI 3.4 or later. The guide calls for three workers with 8 CPUs and 32 GB of memory each, including one GPU node, although it also supplies CPU-only overrides.

What platform teams should inspect

The quickstart is unusually explicit about the difference between a working demonstration and a production design. Its default PostgreSQL and Redis instances are single-replica services intended for exploration; the chart can instead target external services. It includes a sample Argo CD application and supports a SealedSecret path for GitOps-managed credentials.

The default install generates database, cache, JWT and demo-login credentials, creates OpenShift routes and registers the Granite model with LlamaStack. A pinned Aug. 23 revision also removed the registry pull-secret requirement after the application images became anonymously pullable.

Two cautions matter. Red Hat labels the catalog entry as expert-authored but not tested on every supported configuration. The Helm chart and configuration are open source, while the Codvo application images are proprietary. Teams evaluating the pattern should therefore separate the portable architecture—queued agent execution, cited evidence, model serving and auditable state—from the supplied application binaries.

What to try

Run the bundled smoke test after installation, then process one of the sample leases in production mode rather than the synthetic demo mode. Watch the sequential agent pipeline, inspect the evidence pack and compare the final recommendation with its cited clauses. That exercise will expose the operational questions that matter beyond this use case: retry behavior, model latency, state durability, human review and the boundary between generated analysis and an accountable business decision.