Reference implementations of real-world AI-agent workflows, shipped as baseline (plain Python, no Declaw) and sandboxed (running inside a Declaw Firecracker microVM) pairs — so you can see exactly how the same agent hardens end-to-end against the failure modes agents actually hit: PII/PHI leakage, prompt injection from untrusted documents and tool output, tool abuse, data exfiltration, and over-reliance on a non-deterministic model for a material decision.
Every workflow makes real LLM calls (and, where relevant, hits real public APIs). All input data is synthetic.
| Directory | Domain | Workflows |
|---|---|---|
fintech-workflows/ |
Fintech | underwriting, KYC, AML/SAR drafting, chargebacks, compliance RAG, robo-advisor, collections, merchant onboarding, market-abuse surveillance, insurance triage, equity research, tax, treasury, support, fraud explainer, risk narrative |
health-tech/ |
Health-tech | prior authorization, trial matching, medical coding, lab-result explainer, medication safety, MSL literature |
data-intelligence-workflows/ |
Data intelligence | KPI Q&A, telemetry fusion, proactive alerting, chat session |
Each vertical has a workflows/ (baseline) and sandboxed/ (with Declaw) tree of
the same agents, plus its own README.md / SANDBOXED.md / SECURITY.md.
Built with four agent frameworks — LangGraph, CrewAI, AutoGen, and LlamaIndex — to show the controls are framework-agnostic.
The core principle across every demo: the LLM is never the entity that emits a
binding, customer-impacting decision — a deterministic rule or a human gate
owns it, and Declaw makes autonomous execution impossible and audits it. See
GOVERNANCE.md for the convergent core + per-jurisdiction overlays,
and COMPLIANCE-REVIEW.md for the engineering-fact vs.
legal-claim split.
⚠️ These examples are engineering reference material, not legal advice, and use synthetic data only. Validate any regulatory claim with qualified counsel for your jurisdiction before relying on it.
# 1. Install the Declaw SDK + a framework (per workflow requirements.txt)
pip install declaw langgraph openai # example
# 2. Set your keys
export DECLAW_API_KEY=dcl_...
export OPENAI_API_KEY=sk-... ANTHROPIC_API_KEY=sk-ant-... # as needed
# 3. Run a baseline (no Declaw) and its sandboxed counterpart
python fintech-workflows/workflows/01-credit-underwriting-langgraph/run.py
python fintech-workflows/sandboxed/01-credit-underwriting-langgraph/run.pyGet a Declaw API key and read the docs at declaw.ai · docs.declaw.ai.
Apache License 2.0. Copyright 2026 Declaw.