The role
Job description
Must:
LangGraph - design compiled multi-agent graphs: nodes, conditional edges, typed state schemas with correct reducer functions, intent routing, deterministic workflow spines
Human-in-the-loop design - interrupt and resume with HANA-persisted checkpoints; keep the human in control of consequential actions
Understanding Transformer Architecture - Attention mechanisms and enforced Tool-Calling
A2A protocol - implement agent servers: agent card, skill advertisement, streaming JSON-RPC over SSE
MCP - design and host tool surfaces (e.g. via @cap-js/mcp)
Prompt engineering - output format constraints, PII leakage prevention, domain scoping; test under adversarial input, not just happy paths
Know where LLM autonomy ends and deterministic code must take over - and have learned this from production failures, not tutorials
Apply a cost/benefit test before building an agent - a deterministic pipeline with one LLM call is often better
Semantic Search for context ingestion
MLflow - instrument agent runs, log metrics and token usage per turn, compare model and prompt versions across releases
Vector database experience - RAG pipelines and semantic search (HANA Cloud Vector Engine or similar)
Async Python 3.13 - async/await, ContextVars, asyncio.to_thread for blocking drivers; know what stalls an event loop and how to fix it pytest + pytest-asyncio - test suites including async patterns; coverage discipline
python backend development with e.g. starlet or fastapi
Azure DevOps / GitHub Actions - CI/CD pipeline maintenance and troubleshooting
Good to have
AI observability and experiment tracking
Build evaluation datasets; run LLM-as-judge or golden-dataset assertion checks
OpenTelemetry - distributed traces across agent turns and downstream services; OTLP export to SAP BTP Cloud Logging
Dynatrace / Kibana - triage latency spikes and errors by agent node
Token usage tracking per model - cost-control signal, not a curiosity
SAP AI Core, Launchpad, SDK, Generative AI Hub
ruff + mypy + pre-commit - linting, type checking, hook enforcement
N8N - low-code workflow automation for rapid agentic prototyping
EU AI Act and AI ethics - risk classification, transparency requirements, responsible AI patterns for enterprise agents