Case study / ClawKeeper
ClawKeeper, 110 agents inside a deterministic finance hierarchy.
A production AI system for finance ops. Specialized agent roles, policy engine in front, Postgres Row Level Security at the data layer. Every action is auditable.
02 / Problem
Finance teams cannot ship 'agentic' AI on vibes.
Auditability, separation of duties, and approval gates are not optional. Most agent demos collapse the moment a regulator asks who approved what.
03 / Approach
Specialized agents inside a hard policy boundary.
Role hierarchy
Analyst, controller, auditor, approver. Each agent has scope, tools, and an upstream reviewer.
Policy engine
Deterministic rules in front of every action with money or PII. The LLM proposes. The policy decides.
Postgres RLS
Row Level Security enforces tenant isolation and role-level data access at the database, not the app.
Audit log
Every decision, prompt, tool call, and policy verdict captured. Replayable, exportable.
04 / Outcome
Agents finance teams actually trust.
Production deployment with full audit trail. Stack pattern reused across other regulated AI builds. Public source on GitHub.
FAQ
Common questions.
- Why 110 agents?
- Finance work decomposes naturally into specialized roles, analyst, controller, auditor, approver. Each role is a small agent with a narrow scope and a clear audit trail.
- How do you prevent agent drift?
- Deterministic policy engine in front of every privileged action. Postgres Row Level Security enforces what each agent can read or write.
- Is this open source?
- Source is on GitHub. The deployment infrastructure is engagement-specific.
Need agents that survive an audit?
Start with the $1,000 AI Production Readiness Audit. We map your compliance constraints, then design agents around them.