Use cases
5 case-study deep dives
Each of these is a 400-600 word walkthrough of a specific high-leverage shape — the problem, the YAML, what you'd see after 5 real runs, and how it integrates with the tools your team already uses.
- Multi-dimensional PR review — fan-out reviewers + advisor evolution
- Executable on-call runbook — replace wiki rot with versioned, replayable triage
- Multi-vendor compliance scan — Claude + Codex cross-check before release
- Reproducible LLM evaluation suite — eval checkpoints with full trace
- Tier-1 support routing — classify → KB search → auto-reply or escalate
The role-organized index below lists every scenario; deep dives expand the most-built ones.
OpenExpertise isn't general-purpose AI infrastructure. It's specifically the right tool when:
- The same multi-step process runs over and over (every Monday, every PR, every incident).
- The process mixes deterministic logic with LLM judgment (parse + classify, fetch + summarize, lint + review).
- You want the same outcome every time — and a trail showing why.
If that's you, here are the scenarios where the project pays off the fastest, organized by role.
Engineering leads
PR review with multiple specialized reviewers
You want every PR scanned by a bugs reviewer, a perf reviewer, and a tests reviewer — and missed dimensions added by the advisor over time.
→ examples/review-branch — the canonical demo. Three reviewers fan out, find issues, miss SQL injection; the advisor proposes adding a security dimension; re-run catches it.
Why OE wins here: hand-coding this as a serial loop costs a day. As a "let Claude Code review" call, you get inconsistent depth and zero institutional memory. As an OE flow, every PR gets the same triple-pass + the system learns missed dimensions.
Release-gate enforcement
License check + changelog check + coverage check + Claude-Code security scan, all required to pass before merge.
→ examples/release-gates — gate orchestration with mixed tool (license, coverage) + cli-agent (Claude security scan) + agent (final gate decision). 8 minutes to author, runs on every release branch.
Issue triage with dedup
New issue → classify → search similar → if dupes exist, route the dedup decision through an LLM; otherwise auto-assign labels and route to the right team.
→ examples/issue-triage — demonstrates the when: edge for conditional dedup with skip-cascade-breaking via direct edges.
SRE / ops
Incident triage runbook
Page fires → fetch logs from 3 sources → run a parallel investigation across latency, errors, and saturation dimensions → cross-reference findings → assign severity → notify oncall.
→ examples/oncall-runbook — for_each fan-out across dimensions with concurrency: 3. Runs in ~45s, leaves a JSONL trail for the post-mortem.
Why this is hard without OE: incident runbooks usually rot. Wikis go stale; the on-call engineer rediscovers the same dashboards every time. OE makes the runbook executable, version-controlled, and instrumented.
Pre-deploy gate
License compliance + change manifest + smoke test results + LLM-summarized risk → human-readable verdict file.
→ Variant of examples/release-gates — same shape, can be re-pointed at deploy artifacts.
Capacity / cost report
Scheduled oe run against your billing API → aggregate → LLM-written exec summary → POST to Slack.
→ Pattern from examples/dataset-aggregate + an agent summary node + a tool Slack-poster. Wire as a cron job.
Security teams
Compliance scan with structured reporting
Run a Claude-Code-driven security audit + a Codex-driven cross-check + an Anthropic-API verdict synthesizer. Each finding has cite, severity, and remediation in a JSON shape that feeds straight into your ticket system.
→ Variant of examples/tri-cli-orchestration — but with security prompts and a structured-output agent that consolidates the cross-vendor findings.
Why two vendors: cross-checking reduces single-model false positives + false negatives. The 30% cost overhead is worth it for findings that gate releases.
Audit log of every AI-touched decision
Every node touching an LLM emits node.tokens + node.activity events to JSONL. Pipe the JSONL into your SIEM. Now every AI-influenced decision in your org has a per-token trail.
→ See Observability for the integration recipe (Datadog / Splunk / Elastic).
ML platform / AI engineering
Standardized eval suite as YAML
Define an eval graph: dataset → for_each agent (the model under test) → for_each agent (the judge) → aggregate metrics → persist. Same DAG runs for every model checkpoint.
→ Pattern from examples/dataset-aggregate + fan-out agent. Reproducibility is the win: every eval run lives in .openexpertise/runs/ with full token + prompt trace.
Multi-vendor cross-judging
Three judges (Claude, Codex, Gemini) score the same output. Disagreement → flag for human review. No other framework gives you cross-vendor orchestration this cleanly.
→ examples/tri-cli-orchestration is the structural template.
Cost-aware retry policy
LLM call fails on the expensive model → fallback to a cheaper model with on_error: retry + provider-swap in the prompt args. Cost visibility comes from the node.tokens events.
→ See error policies + the run-with-llm guide.
DevRel / Documentation teams
Continuous doc-from-source
Source files change → trigger an OE run → load the relevant section → an agent summarizes the changes → write a draft changelog entry → human reviews. Better than npm-version because the summary captures intent, not just file paths.
→ Pattern: examples/dataset-aggregate + an agent summarizer + a tool writer. Wire to GitHub Actions on push.
Repo audit ("does our README still match reality?")
Compare README.md claims to package.json + actual source files + actual tests. LLM verifies. Surface drift.
→ Build with tool (read files) → agent (compare claims) → tool (write audit report). Run weekly via cron.
Data engineering
Row-level enrichment with model-based classification
Load 10k rows → fan-out an LLM classifier (for_each with concurrency: 16) → aggregate stats → persist. Same shape every Monday morning.
→ examples/dataset-aggregate for the structural template. Swap the dataset source to your warehouse via source: { type: sqlite, ... } or write a small tool that wraps your warehouse client.
LLM-augmented ETL
The "T" step in ETL gets an LLM column (entity resolution, sentiment, classification) without rewriting your existing pipeline.
→ Wire OE as the agent layer your existing airflow / dagster task calls via oe run subprocess. See Deployment.
Product / customer ops
Onboarding workflow with quality gates
New customer signs up → fetch their org data → run a dataset node loading from your CRM → an agent writes a personalized welcome plan → another agent reviews it for tone + accuracy → if approved, tool posts to Slack.
→ Pattern: tool → dataset → agent → agent (reviewer) → conditional tool. See cookbook fan-out + conditional skip.
Tier-1 support routing
Incoming ticket → classify (urgency × topic) → search KB → if duplicate, auto-reply with link; else route to the right human queue. The "if duplicate" branch uses an LLM to write the response in the customer's tone.
→ Direct adaptation of examples/issue-triage — swap GitHub issue source for your ticketing source.
Knowledge work meta-loops
Three superpowers skills already shipped as OE flows
The superpowers skills repository contains battle-tested cognitive workflows. We've ported the three most evolution-friendly ones to runnable OE experiences:
brainstorming— diverge across 3 angles → cluster → critique → synthesize top 3systematic-debugging— hypothesize → verify → fix → re-testdeep-research— multi-source research with cross-vendor cross-referencing
These are open invitations: take a skill that works for your team, port the steps to YAML, let evolution refine it.
When OpenExpertise is not the answer
| Scenario | Use this instead |
|---|---|
| One-shot exploration ("help me understand X") | Claude Code / Cursor / Codex / Gemini direct |
| You want an autonomous agent that reasons freely | Claude Code / Codex / Gemini direct |
| Real-time chat assistant | Vercel AI SDK / OpenAI Assistants / Anthropic Claude |
| Embedding / vector DB pipelines (no agentic step) | LangChain / LlamaIndex |
| Pure deterministic data pipeline (no LLM) | Dagster / Airflow / Prefect |
| You need durable execution > 24 hr with retries on every op | Inngest / Temporal (or layer OE on top of them) |
OpenExpertise sits above Claude Code / Codex / Gemini, not beside them. It sits adjacent to LangGraph / CrewAI / Mastra (different design, similar problem space). See Compare for the full matrix.
→ Start with the example closest to your shape, then bend it to your data. The cookbook has the most-reused patterns extracted.