AgentForge is domain-agnostic — the platform is the agent; an example is just a corpus + a system prompt + config. Each one below runs on the same code.
| Example | Domain | Run |
|---|---|---|
banking-compliance |
AML/KYC policy + product sheets, with HITL before filing a SAR / escalating a case | default (docker compose up) |
insurance-support |
Auto-insurance policy, claims process, fraud escalation | docker compose -f docker-compose.yml -f docker-compose.insurance.yml up |
- Drop your
.mdcorpus underexamples/<name>/corpus/. - Point ingestion at it (
AUTO_INGEST_CORPUS) and pick aCOLLECTION_NAME. - Set
SYSTEM_PROMPTto frame the assistant for your domain (keep the "answer only from context, cite, refuse when out of scope" rules). - Optionally add domain-specific tools in
agentforge/agents/tools.pyand list the sensitive ones inSENSITIVE_TOOLSso they require human approval. - Add an
evals.jsonland gate on it in CI.
The insurance example is the smallest possible diff that does all of the above —
copy its config.yaml and compose overlay as a template.