MiMi Money AgentOS

Inspiration

MiMi Money AgentOS was inspired by MiMi Money with the need to overcome Africa's fragmented financial ecosystem by using agentic AI to create an affordable, autonomous digital workforce that can operate financial services intelligently, securely, and at scale in MiMi Money.

MiMi Money is a Super mobile app that combines digital wallet financial functionality, communication, Web3, AI agents, and social and business networking in one platform. MiMi Money is already live in Google Play Store and a contender in Google's Build with Gemini Xprize hackathon.

The inspiration for MiMi Money AgentOS came from seeing how fragmented financial services still make everyday digital finance difficult across Africa. People and businesses often have to interact with multiple wallets, payment platforms, banks, exchanges, and financial applications that do not work together seamlessly. At the same time, financial institutions and fintechs rely heavily on human teams to monitor transactions, investigate fraud, perform compliance checks, reconcile payments, and respond to customers.

We were inspired by the emergence of agentic AI and the possibility of moving beyond AI that simply answers questions to AI that can plan, collaborate, make decisions, use tools, and complete work autonomously. Rather than building another financial chatbot, the idea was to create an AI workforce for financial operations where specialized agents such as Payment, Fraud, Compliance, Risk, Reconciliation, Treasury, and Customer Support agents work together like employees in a digital organization.

Another major inspiration was the ease and speed of developing sophisticated software with modern AI coding tools. AI makes it increasingly possible for a small team or even a single builder to create systems that previously required large engineering and operations teams. This creates an opportunity to build an affordable financial infrastructure layer specifically suited to African markets.

MiMi Money, Agenticous AI agent and now MiMi Money AgentOS are operated by a legally incorporated company GETO TECHNOLOGIES - SMC LIMITED, SMC stands for Single Member Company where the rest of the workforce are and will be AI agents.

MiMi Money AgentOS therefore brings these ideas together: AI agents + financial infrastructure + autonomous orchestration + governance, creating a system where routine financial operations can happen automatically while sensitive decisions remain protected by policies and human oversight.

What it does

MiMi Money AgentOS is an autonomous AI financial operating system to provide intelligent financial and operational services to MiMi Money, the mobile application that combines digital wallet functionality, communication, Web3, AI agents, and social and business networking in one platform.

MiMi Money AgentOS is essentially the AI workforce behind MiMi Money.

MiMi Money in Action

MiMi Money AgentOS runs a 10-agent financial workforce Orchestrator, Payment, Fraud, Compliance, Risk, Reconciliation, Treasury, Support, Reporting, Human Escalation organized into departments with real reporting lines, distinct permissions and budgets. Reasoning is powered by Gemini AI specifically Gemini 3.6 Flash through Google ADK agents.

Processes USDC/USDT along African local currencies payments end-to-end: a payment event flows through Pub/Sub → Payment Agent → Fraud ∥ Compliance → Risk → Treasury → Reconciliation → customer notification → audit report all real backend state, no animations.

Halts for humans when policy says so: A payment makes the Fraud Agent score 0.90 risk; a deterministic Policy Engine (outside the LLM) blocks autonomous completion, creates a real Human Review task, and resumes the exact workflow from persisted state only after an operator clicks Approve.

Keeps an inspector-grade UI: live workflow graph, event stream, Memory Bank (Firestore), audit log, policy center, budgets everything driven by actual events

How we built it

We built MiMi Money AgentOS by combining AI agents, financial APIs, blockchain infrastructure, and an event-driven orchestration layer into one autonomous financial operating system.

We used AI-assisted development to rapidly design, code, test, and iterate the platform, while specialized agents handled tasks such as customer support, transaction monitoring, financial operations, and decision-making. The system was designed around an observe → analyze → reason → verify → execute → audit loop, allowing agents to act autonomously while applying security and transaction controls before execution.

We connected the agents to digital payment infrastructure, wallets, stablecoins, blockchain networks, and external services through APIs, while maintaining structured logs and audit trails for every important action. This approach allowed us to build a system capable of operating complex financial workflows with minimal human intervention, demonstrating how AI agents can reduce operational complexity and make advanced financial services more accessible across Africa.

It was built in six layers, extending the existing Paperclip architecture rather than replacing it:

1. Forked the architecture, not just the brand. Paperclip's Express server, React board UI, org-chart/approval/budget/activity/heartbeat systems were preserved intact and reused as the enterprise control plane; the financial domain was added as a clean extension.

2. packages/mimi (@mimi/agentos-core) — a framework-free TypeScript engine implementing the whole domain: agent catalog & org chart (10 agents, departments, reporting lines), the deterministic Policy Engine (thresholds, fraud bands, sanctions, rate limits, budget hard-stops), the MiMi Agent Gateway (identity → RBAC → policy → human approval → verified tool → audit), a Pub/Sub-semantics event bus with idempotency/retries/DLQ (+ emulator/production REST mirror), a Firestore-shaped Memory Bank (+ Firestore REST client), 12 structured financial tools, the crash-safe persisted-state payment orchestrator, and African-market seed data.

1. Three interchangeable runtimes behind one AgentRuntimeAdapter contract: GoogleAdkRemoteRuntime (delegates to the real google-adk Python service), GoogleAdkGeminiRuntime (in-process Gemini function-calling loop), and GoogleAdkLocalSimulationRuntime (deterministic offline mode for zero-cost dev/CI).

2. services/mimi-adk-runtime — the workforce as real Google ADK Python agents (google.adk.agents.LlmAgent + Gemini), exposed via FastAPI (/healthz, /runs). Its tools call privileged capability back through the Gateway HTTP surface, so security enforcement is identical no matter where reasoning happens.

1. Server + UI integration: the MiMi router mounts under /api/mimi inside the existing Paperclip app (with activity mirroring into Paperclip's log), and the board UI gained 13 MiMi pages (Overview, Fleet, Workflows graph, Reviews, Memory Bank, Audit…) built with the existing component system and a token-gated dark theme, updating live over SSE from real backend events.

2. Verification-first engineering: 46 core tests (policy bands, gateway denials, event idempotency, crash recovery, retry-then-complete, no chain-of-thought persistence), 7 server E2E route tests, Python unit tests for the ADK service, plus full workspace typecheck/build/token-gate gates — both demo flows were validated end-to-end over live HTTP before hand-off.

Challenges we ran into

Developing MiMi Money AgentOS presented several technical challenges because we were not simply building a chatbot, we were building a distributed AI financial workforce where multiple autonomous agents needed to collaborate reliably, securely, and asynchronously.

1. Multi-agent orchestration — Coordinating Payment, Fraud, Compliance, Risk, Reconciliation, Treasury, and Support agents was more complex than running a single AI model. We had to design clear responsibilities, communication protocols, dependencies, retries, and handoffs between agents.

2. Making AI agents truly autonomous — Getting Gemini to produce an answer was relatively straightforward; getting agents to reliably plan, use tools, execute multi-step workflows, verify their actions, and recover from failures required considerably more engineering.

3. Agent-to-agent communication — We needed a reliable mechanism for agents to communicate without creating tightly coupled services. Pub/Sub became important for asynchronous events, but this introduced additional challenges around event ordering, duplicate messages, retries, dead-letter queues, and idempotency.

4. Financial security and permissions — An AI agent cannot be allowed to freely execute financial operations. We had to build an Agent Gateway and permission model that determines exactly which agent can access which tool and what actions it is allowed to perform.

5. Preventing unsafe AI decisions — LLMs can make incorrect or unexpected decisions, so critical financial policies had to remain deterministic and outside the model. For example, a high-value or high-risk transaction must automatically trigger human approval regardless of what Gemini recommends.

6. Persistent agent memory — Maintaining useful context across long-running workflows was challenging. Firestore had to store agent state, transaction context, workflow progress, previous decisions, and operational summaries without exposing sensitive information or private model reasoning.

7. Long-running asynchronous workflows — Financial operations may take minutes or longer and may involve several agents. Designing workflows that could pause, wait for human approval, resume, retry, or recover after an agent failure required persistent state management.

8. Real-time AgentOS UI — Making the Paperclip-based MiMi AgentOS interface reflect what was actually happening in the backend was challenging. Agent statuses, workflow graphs, events, approvals, and audit logs needed to be driven by real backend events rather than simulated animations.

9. Integrating Paperclip with Google ADK — Paperclip serves as the orchestration/control plane while ADK handles the actual AI agents. Connecting the two cleanly required an adapter layer so that agent lifecycle, tasks, heartbeats, approvals, and execution state could flow between the systems.

10. Observability and debugging — When several AI agents, Pub/Sub events, Cloud Run services, tools, and databases are involved, finding the source of a failure becomes difficult. We needed correlation IDs, workflow IDs, agent run IDs, structured events, and audit logs to trace operations end-to-end.

11. Cloud deployment and cost control — Running an agentic system entirely in the cloud can become expensive. We had to design the architecture around Cloud Run, Pub/Sub, Firestore and Gemini while minimizing always-on infrastructure and unnecessary model calls.

12. Handling failures and recovery — Gemini can fail, an agent can timeout, a tool can return an error, or a Pub/Sub message can be delivered more than once. We therefore had to design retry, timeout, idempotency, checkpointing, and recovery mechanisms so a financial workflow would not accidentally execute an operation twice.

13. Testing autonomous behavior — Traditional unit tests are not enough for multi-agent systems. We needed to test complete workflows, including low-risk transactions, fraudulent transactions, policy violations, human escalation, agent failures, duplicate events, and workflow recovery.

14. Maintaining security without destroying autonomy — One of the hardest design problems was finding the balance between allowing agents to work independently and ensuring they cannot bypass financial controls. The solution was to give agents bounded autonomy: agents can act independently within predefined capabilities and policies, while exceptional actions require human intervention.

The biggest technical challenge was turning multiple probabilistic AI agents into a reliable, secure and observable financial workforce capable of operating autonomously while ensuring deterministic policies, persistent state, human approvals and auditability remain in control.

Accomplishments that we're proud of

We are proud to have transformed the concept of combining AI agents into MiMi Money AgentOS, a governed digital workforce where specialized AI agents can collaborate autonomously on financial operations while remaining observable, auditable, secure, and under human control when it matters.

What we learned

Developing MiMi Money AgentOS taught us that AI agents can fundamentally change how financial services are built and operated, especially in markets where traditional financial infrastructure is fragmented. We learned that AI can replace many repetitive operational tasks, allowing a small team or even a single developer to build, monitor, support, and operate sophisticated financial workflows without a large workforce.

We also learned that building agentic financial systems requires more than intelligence: security, transaction verification, human oversight, auditability, reliability, and clear guardrails must be built into the architecture from the beginning. Most importantly, we learned that combining AI agents with digital payments, stablecoins, automation, and interoperable financial services can create a more accessible financial platform for Africans, while AI-assisted development dramatically reduces the time, cost, and complexity required to turn these ideas into working products.

What's next for MiMi Money AgentOS

MiMi Money AgentOS has been the missing agentic AI layer in MiMi Money to power the financial services of the AI Agents, Pay and Bank buttons in MiMi Money for an AI workforce for financial operations where specialized agents for Payment, Fraud, Compliance, Risk, Reconciliation, Treasury, and Customer Support agents work in the background together like back office employees in a digital organization.

As we rollout MiMi Money AgentOS the next step for MiMi Money AgentOS could be to evolve into a fully autonomous financial operating system demo model AgentOS for startups and established financial businesses. We could expand the agent ecosystem with specialized agents for payments, customer support, fraud and transaction monitoring, treasury management, lending, compliance, accounting, and business operations in MiMi Money or a service to other African financial institutions struggling with the same problems we are solving.

MiMi Money AgentOS could also connect more deeply with mobile money, banks, stablecoins, blockchains, and fintech APIs across Africa, allowing users to manage fragmented financial services through a single intelligent interface. Over time, the platform could support multi-agent collaboration, where hundreds of specialized AI agents coordinate tasks autonomously while maintaining strict security, permissions, audit trails, and human oversight. The ultimate vision would be to make sophisticated financial infrastructure accessibly, affordably run with MiMi Money AgentOS and turning MiMi Money AgentOS into the autonomous financial operating layer for MiMi Money and other African financial institutions in Africa's digital economy.

Built With

• cloud-run
• firestore
• gemini-3.5-flash
• gemini-3.6-flash
• google-adk
• google-cloud
• google-gemini
• pub/sub