Manila Before Manhattan

Three levels of financial autonomy and the case for building in frontier markets first

It's the early 1600s in Manila, and the Catholic Church has a problem most institutions would envy: too much idle capital. Their solution looks quaint in 2026: become venture capitalists.

Wealthy Spanish colonists preparing for the trans-Pacific galleon crossing faced a voyage so dangerous that writing a will beforehand was standard practice. They bequeathed fortunes to the Church's charitable endowments, the Obras Pias. The money was meant for orphanages, and hospitals. Noble purposes, all of them. But the friars who managed these funds recognized something that any modern fund manager would: cash sitting in a vault compounds at zero.

So they did what any rational capital allocator would do. They lent it out. Funds for charity morphed into tools of trade finance.

The borrowers were merchants sailing the Manila-Acapulco Galleon Trade, and the terms reflected the risk. If your ship made it to Acapulco and back, escaping typhoons, pirates, and the Dutch navy, the Obras Pias collected enormous returns. If your ship sank, the loan was written off. This was literally venture capital with a splash of commercial lending and maritime insurance in colonial Manila.

For over two centuries, the speculative profits of this arrangement funded the physical construction of the colonial Philippines: the stone churches, the roads, the first institutional architecture of a financial system that hadn't existed before the trade created the need for one.

The galleon trade came first. The institutions followed.

The pattern repeated in the 1800s, after the galleon trade collapsed and Mexican independence cut off the colonial subsidy that had kept the Philippines fiscally afloat for three centuries. The archipelago was forced to become a producer. Manila's ports opened to foreign trade. Sugar, abaca, tobacco, and coffee boomed.

But the Obras Pias couldn't finance industrial-scale agriculture. They were built for a different era. Into the void stepped British and American merchant houses. Nicholas Loney, the British Vice-Consul in Iloilo (my father's hometown, where my grandmother was the first chief librarian of the American period), imported steam-powered sugar mills from Scotland on installment plans and extended crop loans to local planters. Loney single-handedly transformed Negros Island into the sugar capital of Southeast Asia.

HSBC arrived in 1875 and financed the country's first railway. El Banco Español-Filipino, seeded with Obras Pias capital, began issuing the archipelago's first paper currency. It exists today as the Bank of the Philippine Islands - the country's 3rd largest bank.

Within a generation, the profitable commodity trade had bootstrapped an entirely new financial system. Formal banks, paper currency, railways, telegraph networks became the infrastructure that the colonial government had failed to build in three hundred years of direct rule.

This is how economic systems are actually born. Not by decree, and rarely by central planning. A new form of commerce emerges. Early participants capture outsized returns. Those returns get levered into building the infrastructure the new commerce requires. And that infrastructure, once built, enables orders of magnitude more activity than what originally funded it.

Light bulbs were invented before the grid. Planes before airports. Galleon profits before the banking system. Crop loans before the railway.

There are good reasons to believe AI agents in finance are entering this same cycle. Agents that manage cash positions, chase invoices, reconcile books, and negotiate vendor contracts are already creating real economic value. Agents are collapsing costs, recovering revenue, replacing operational capacity that simply doesn't exist in most emerging markets.

Each deployment generates returns that fund the next layer of infrastructure: agent-native accounts, settlement protocols, machine credit, programmatic compliance. Each layer unlocks use cases that weren't possible before it existed.

But the lessons of history face their own friction in the present.

The galleon merchants didn't need a banking license. Loney didn't need regulatory approval to sell sugar mills on credit. They operated in a vacuum of institutional authority, and they moved fast because no one was in a position to stop them.

Financial agent infrastructure doesn't have that luxury. It operates inside the most heavily regulated sector in every economy on earth. What gets built first, what requires bank charters and licenses and franchises… the entire sequencing of this infrastructure buildout is the entire game.

This essay lays out that sequence. It's a useful guide to builders and policymakers on how to rationally think about the opportunities associated with agentic finance.

Agentic finance leapfrogs hardest where traditional financial infrastructure is weakest. The world saw this with mobile phones leapfrogging landlines. The same logic applies here — brittle institutional infrastructure doesn't just mean agents optimize existing workflows. It means they replace infrastructure that never existed.

We've organized what follows into three levels, borrowing the autonomy framework from self-driving vehicles.

Level 1 covers what's achievable today with current AI capabilities.

Level 2 describes what becomes possible in 2027-2028 as multi-agent coordination and reasoning improve.

Level 3 looks further out — to a world that requires regulatory shifts and institutional agreements that don't yet exist.

The timeline may turn out to be conservative. But the sequencing matters more than the speed.

It goes without saying that this is not a comprehensive list. We look forward to an abundance of use cases yet to be known.

Level 1 Agents: Automated Workflows

In the Philippines, most SMEs run finance on spreadsheets and group chats. Invoice follow-up happens over Viber. Bill pay is one-by-one through bank portals. Expense reimbursements are paper-based. Reconciliation is manual.

And the people who could professionalize these operations are disappearing. Accounting student enrollment has declined 41% and CPA board examinees have dropped 35% since 2019.

Level 1 agents solve this with capabilities that exist today. Current AI frameworks can already use tools, produce structured output, and integrate with banking APIs to read balances, trigger payments, and write to ledgers. Protocols like x402 allow agents to hold funds and pay for data and services over HTTP — meaning the plumbing for agent-initiated transactions is already live.

The scenarios below could all be built and deployed now.

BIR Compliance Agent

Ana, owner of a 15-person logistics company in Cebu

Problem: Ana pays an accountant ₱30,000/month to handle BIR tax compliance: preparing 2307 withholding certificates, filing quarterly returns, reconciling VAT input and output across her PHP and USD accounts. The accountant makes errors that Ana only discovers when she gets an assessment notice.

Agent Solution: An agent monitors all transactions, auto-classifies them by BIR tax category, generates 2307 forms for each vendor payment, and prepares quarterly filings in the format BIR's eFPS system expects. It flags transactions it can't confidently classify and routes them to Ana with a one-tap approve/edit interface. At month-end, it reconciles her peso and dollar accounts across two banks and produces a single consolidated view.

Impact: Ana's compliance costs drop from ₱30,000/month to near zero. She files on time every quarter without hiring a full-time accountant she can't afford — and that the market can't supply.

Invoice Agent

Tomoko, freelance design studio owner with 8 active clients

Problem: Tomoko sends invoices through her neobank but has no systematic follow-up. She realizes invoices are overdue only when cash gets tight, then awkwardly emails clients one at a time.

Agent Solution: An agent monitors all outstanding invoices. At 7 days past due, it drafts a reminder email. At 14 days, the tone firms up. At 30 days, it flags the invoice for Tomoko with a recommended action. Tomoko approves templates once; the agent handles the rest. Past 30 days, the agent autonomously offers a collections notice, working with a loan service agent maintained by another startup, or offer invoice financing to help smoothen the cash gap.

Impact: Average days-to-payment drops from 45 to 19. Tomoko never has to write an awkward collections email.

Accounting Reconciliation Agent

Raj, startup controller managing a QuickBooks integration

Problem: Every month-end, Raj spends two days reconciling bank transactions with QuickBooks. The sync misses edge cases: refunds coded wrong, transfers between accounts double-counted, international wire fees categorized inconsistently.

Agent Solution: An agent runs nightly, pulling all new transactions and comparing them against QuickBooks entries. It identifies mismatches, suggests corrections, and auto-fixes known patterns (e.g., always categorize bank wire fees as 'Bank Service Charges'). Raj gets a morning report of what was auto-fixed and what needs his judgment.

Result: Month-end close drops from 3 days to 4 hours. The error rate on categorization falls to near zero.

Level 2 Agents: The Autonomous Corp

This level requires multi-step reasoning and cross-system orchestration across disconnected systems (say, your operating bank and your asset manager), and richer agent-to-agent coordination.

These scenarios depend on capabilities actively being developed: multi-agent orchestration, long-horizon planning, real-time tool use across multiple APIs simultaneously, and more reliable autonomous decision-making under uncertainty.

Level 2 requires banks to support agent-native account types.

Level 2 is when agentic finance becomes truly multiplayer (vs single player workflows in Level 1). If Level 1 involved agents initiating transactions on behalf of a human (and at most collaborating with other agents), Level 2 will see agents transacting with each other.

One interesting observation is that overseas diaspora remittances are an existing agent-to-agent payment flow hiding in plain sight.

For example, ~$40B/year flows into the Philippines from overseas workers, and the chain is this: Overseas Filipino worker earns abroad → sends via remittance app → family receives in local wallet → pays bills/tuition/rent → sends confirmation back to the sender that funds have been used → asks for more the following month. This is a multi-hop, multi-party automated flow with terrible economics (3-5% fees, 1-2 day settlement).

Is this a Level 2 finance agent scenario made specific? An agent-native remittance rail can settle via stablecoins and auto-distribute to billers on the receiving end sounds like an obvious product.

Beyond families, we explore the following scenarios for businesses below:

Autonomous A/P

No human operator — a fully autonomous AP agent

Problem: A 200-person company processes 400 bills/month across 3 approval levels and 80 vendors. The Accounts Payable team needs to scale alongside the business, but resources are tied up to compute and AI engineers. We've talked to BPO finance teams in Makati who spend 2 to 3 full-time headcount on exactly this workflow.

Agent Solution: An agent network handles the entire AP lifecycle: The Bill Agent receives and parses invoices from email, matches to POs from the procurement system, and codes to GL accounts. The Approval Agent routes to the right approver chain based on amount and department, follows up on stale approvals, and escalates blockers. Approvers (like function or business unit heads) are represented by Policy Agents with built-in fiduciary and governance skills to make approve / decline decisions on behalf of their principals. The Payment Agent queues approved bills for payment on optimal dates, handles international wire requirements for overseas vendors, and archives payment confirmations. Compliance Agent ensures every payment has proper documentation, flags unusual vendor banking changes, and maintains the approved vendor list.

Impact: The company runs AP with zero full-time staff. The finance lead reviews a daily summary dashboard. Exception rate: under 1%.

Vendor Negotiation Agents

Javier, Operations Manager at a scaling startup

Problem: The company's SaaS spend has grown 3x in a year but nobody has renegotiated any vendor contracts. They're paying list price on tools they've outgrown and still paying for tools nobody uses.

Agent Solution: An agent analyzes 12 months card transactions by vendor: total spend, usage trends, per-seat costs, and contract renewal dates. It identifies the top 10 renegotiation opportunities by potential savings, drafts outreach emails to each vendor with specific asks (volume discount, annual prepay discount, seat reduction), and tracks responses. It knows common discount thresholds for major SaaS tools.

Impact: The company saves $180K/year in SaaS spend through systematic renegotiation it never would have done manually.

Cross-Company Supply Chain Agent

A startup and its key supplier, using an AI-native bank

Problem: The startup needs 60-day payment terms but the supplier needs cash within 15 days. Today this mismatch causes friction, manual negotiations, and sometimes lost suppliers.

Agent Solution: A supply chain finance agent negotiates directly with the supplier's agent. The bank sees both sides of the relationship. Thus, it offers dynamic discounting: the supplier can get paid early at a small discount, funded by the bank's supply chain finance product. The agents negotiate the optimal discount rate based on both parties' cash positions, agree on terms, and execute the payment — all without human intervention beyond initial policy setting.

Impact: Supply chain financing happens automatically between trusted parties, eliminating billions in market inefficiency.

Global Procure-to-Pay Agents

Actors: a manufacturer's procurement agent, supplier agents, logistics agents, agentic bank accounts on all sides

Problem: A vehicle manufacturer orders components from 50 suppliers across 5 countries. Purchase orders, invoices, shipping confirmations, quality inspections, and payments are all separate manual workflows involving 20 different humans.

Agent Solution: The manufacturer's procurement agent issues a machine-readable RFQ via MPP. Supplier agents respond with pricing, lead times, and quality certifications. The procurement agent selects winners, creates purchase orders as smart contracts. An escrow agent facilitates the contracts. When the supplier's logistics agent confirms shipment (verified via a shipping API), partial payment releases. When the startup's receiving agent confirms quality inspection, final payment releases from the escrow agent. All cross-border payments flow through agent-native operating accounts with automatic FX optimization.

Payment Mechanic: International wires and/or stablecoin payments triggered by agent-verified milestones. Escrow accounts per PO. FX locked at commitment, settled at delivery.

Impact: The entire procure-to-pay cycle runs without humans. A 50-supplier, 5-country supply chain settles autonomously.

Level 3: The Machine Economy

Levels 1 and 2 describe what we can build and sell in the next three years. Level 3 points toward something larger: a financial system where agents are first-class economic actors — holding accounts, pricing risk, extending credit to each other, and settling autonomously. Here's an example.

The Financial World Model

An Asian conglomerate's executive runs its annual planning process across 50+ subsidiaries

Problem: Strategic decisions (hire 2000 people? open a new market? acquire a competitor?) require weeks of financial modeling that's outdated the moment it's complete.

Agent Solution: A financial digital twin operates as a world model: it maintains a real-time simulation of the conglomerate's finances on an entity and business-unit level, fed by bank data, payroll, CRM, and operational metrics, along with macroeconomic data from government sources. It's complemented by the company's tribal knowledge, connected through a conglomerate-spanning knowledge graph. Executives can ask natural-language questions: 'What happens if we hire 500 engineers in Q3?' 'Can we afford to acquire Company X at $5B?' 'What's our break-even date if we lose our biggest customer?' The twin runs Monte Carlo simulations in seconds, showing probability distributions of outcomes rather than single-point estimates.

Impact: Strategic financial decisions backed by real-time simulation instead of stale spreadsheets. Decision latency drops from weeks to minutes.

Beyond the financial world model, the full vision of Level 3 includes inter-agent derivatives markets for machine resources, autonomous credit scoring for non-human entities, and self-organizing economic coordination — what you might think of as monetary policy for machine commerce. We're not going to spec these in detail because the honest answer is that nobody knows exactly what form they'll take. What we do know is that each of these capabilities depends on standards and regulatory frameworks being built at Levels 1 and 2.

Getting there requires infrastructure that doesn't exist yet: verifiable agent identity standards, regulatory frameworks that recognize non-human transacting entities, and interoperability agreements among financial institutions.

These are hard problems on the scale of SWIFT or the DTCC, or the early years of Dee Hock building the Visa Network. If levels 1 and 2 are measured in acceleration on a quarterly scale, level 3 is measured in decades. The history of standard setting in technology as described in "Engineering Rules" becomes required reading for all AI professionals.

We think the path runs through emerging markets first. Countries building new financial rails from scratch have less legacy architecture to retrofit and stronger incentives to experiment.

The same leapfrogging logic that brought mobile money to Kenya before Kansas will likely bring agent-native banking to Manila before Manhattan.

We're building Reach Labs to be ready when that happens. But we're not waiting for it. Everything we ship today is designed to work at Level 1 and compound into Level 2.

Reach for the Rails

Each level in this essay describes a deeper layer of financial infrastructure.

Level 1 is plumbing: connecting existing APIs to agent wallets and payment protocols so your agents have funding and data. Level 2 is infrastructure: escrow, settlement, multi-agent accounting, and machine credit so agents can transact with each other. Level 3 is the economic substrate itself: identity, reputation, and monetary policy for machines.

What most writing on agent finance misses is that the political-economic sequence matters more than AI's scaling laws.

In markets with deep, mature banking infrastructure, agentic finance is an optimization layer. It makes existing workflows faster and cheaper. This is valuable, but incremental.

In frontier markets, where 80% of businesses have no finance teams, where cross-border settlement takes days and costs 3-5%, where a single CPA serves dozens of businesses because there aren't enough to go around; agentic finance is first contact with first world infrastructure.

The Philippines has 1.1 million SMEs that need financial operations capacity they can't hire. Cross-border flows are worth $312B, driven by trade, BPO revenue and remittances. This flows through infrastructure built for a pre-agent world. 25% of domestic deposits are in US dollars. It has a regulator in BSP that has already shown willingness to experiment with digital banking licenses and e-money frameworks.

The raw ingredients are there. What needs to be built is the platform layer via AI-native financial infrastructure that lets agents access banking, payments, and credit primitives to operate across these fragmented systems on behalf of the people and businesses who need it most.

Every primitive we ship (programmatic accounts, payment APIs, stablecoin wallets and on/offramps, virtual card issuance, treasury and yield, accounting sync, approval workflows) is designed to work at Level 1 today and compound into Level 2 as agent capabilities mature.

Right now, that agent infrastructure gap is in emerging markets. The teams that build there first will define what AI-native finance looks like everywhere.