AWS Builder Center

Builder Center

Agents for Humans and How We Proved Recoup Works: 419 Backend + 127 E2E Tests and Zero Unsafe Actions

Agent demos are easy to hand-wave; we built ship gates judges can re-run. Building Recoup: Agents for Humans and How We Designed a Safe AWS Recovery Workflow with Strands Graph covers architecture; this post is proof: J-FULL Playwright, golden SLA Strands replay, 419+127 tests, six scorecard gates, zero unsafe actions.

Agent demos are easy to hand-wave, we optimized for ship gates judges can re-run the same bar as our Devpost submission and demo video.
Building Recoup: Agents for Humans and How We Designed a Safe AWS Recovery Workflow with Strands Graph  explains the Strands SLA graph; Agents for Humans — Making AI Financial Decisions Trustworthy: Cedar, Evidence Redaction, and HITL in Recoup explains trust on the live demo. Here we show two complementary proofs:
  1. Operator journey (J-FULL) - scan → three Start Recovery paths → HITL triage → Recovery Ledger (+ SNS on approve), exercised in Playwright and ./scripts/prod_journey_hitl_smoke.sh against production API.
  2. Agent depth (repo + CI) - golden SLA replay through the 11-node graph with Decimal math (never LLM numbers) and adversarial policy/safety tests.

Test pyramid

  • 419 backend tests (pytest), including golden SLA replay and adversarial policy/safety cases
  • 127 Playwright tests across 16 specs, including J-FULL (scan → three triage paths → ledger + SNS) and PSC guest-session isolation
Run locally:
1
2
cd frontend && npx playwright test e2e/journey-full-discovery-triage-ledger.spec.ts
cd backend && pytest -q

Deterministic financial math

SLA credit recovery uses Decimal arithmetic and golden fixtures never LLM-generated numbers. The replay adapter lives in backend/src/recoup/adapters/replay.py; CI enforces P95 latency and consecutive pass counts.

Ship gates

/api/quality/scorecard exposes six gates (golden SLA replay through the Strands graph, financial math, evidence recall, tool autonomy, trace completeness, zero unsafe external actions). CI runs scripts/assert_ship_gates.py.
For demo video Scene 6 run locally (all gates pass when fixtures are present):
1
curl -s http://localhost:8000/api/quality/scorecard | jq '.all_gates_pass'
Production API may be used for smoke checks; golden replay gates on App Runner can fail if fixture paths differ prefer local for “all gates pass” claims.

Operator journey smoke

1
./scripts/prod_journey_hitl_smoke.sh
This script exercises approve / investigate / decline on three distinct services against production API with isolated demo sessions no browser, no AWS keys (matches hackathon “try it” expectations).

Try the live demo

https://pdkeexzwxr.us-east-1.awsapprunner.com/scan  → Demo Scan → Start Recovery on three services → triage on detail → Recovery Ledger. Guest sessions: sidebar Reset Demo Data.
Repo: https://github.com/swa01wk/recoup 
Any opinions in this article are those of the individual author and may not reflect the opinions of AWS.
Enjoyed reading this content? Let the author know!

Your likes, comments, shares, and saves help creators reach more builders.

Loading recommendations

Loading article