
Building Recoup: Agents for Humans and How We Designed a Safe AWS Recovery Workflow with Strands Graph
FinOps tools excel at finding waste; teams still struggle to close recovery safely. Recoup uses an 11-node Strands graph on Amazon Bedrock, with destructive power gated by deterministic Cedar policy and human approval, not LLM execution.
FinOps tools excel at finding waste; teams still struggle to close recovery safely. For the Agents for Humans hackathon we built Recoup, an autonomous cloud-spend recovery agent with an 11-node Strands graph on Amazon Bedrock, but keeps destructive power behind deterministic Cedar policy and human approval.
Graph
Recoup’s pipeline is deliberately split:
- 5 agent nodes - e.g. incident correlation, evidence collection, eligibility reasoning, claim packaging, case monitoring
- 6 deterministic nodes - normalization, SLA math, sanitization, Cedar policy gate, submission adapter, monitoring hooks
The LLM proposes; Cedar decides; humans approve at the boundary.
Flow:
normalize_event → incident_correlation [Strands] → evidence_collection → eligibility → claim_package [Strands] → risk_policy_gate [Cedar] → HITL → submission_adapter → case_monitorSee the canonical diagrams (Mermaid) in our repo: architecture/architecture.md - stack, J-FULL lifecycle, and agent graph.
Why not “LLM executes”?
Financial remediation is high stakes. We enforce:
- Tool allowlists and autonomy classes (GREEN / YELLOW / RED / BLACK)
- Cedar default deny on writes
- Explicit
REQUIRE_APPROVALfor sensitive actions - Claim binding:
claim_hash, amount, andstate_versionmust match on approve (tamper → HTTP 409)
Code pointers:
- Graph nodes:
backend/src/recoup/graph/nodes.py - Cedar policy:
infra/policy/recoup-policy.cedar
Try the live demo
No AWS keys required: https://pdkeexzwxr.us-east-1.awsapprunner.com/scan → Demo Scan → Start Recovery on three services → approve / investigate / decline on opportunity detail → Recovery Ledger. Guest sessions:
X-Demo-Session + sidebar Reset Demo Data.Enjoyed reading this content? Let the author know!
Your likes, comments, shares, and saves help creators reach more builders.
Loading recommendations
Loading article