Your AI Agent Has Nowhere to Land
Today's agents can answer, suggest, even decide. Almost none of them can produce a real business fact — an invoice that hits your books, a stock receipt that updates the warehouse, a payroll entry your accountant trusts.
A modern LLM can write a perfect order in chat. It can negotiate, propose, even argue. But none of that is an order until something — somewhere — generates the invoice, reserves the stock, updates the customer balance, posts the journal entries and leaves a trail your accountant accepts. Without that 'somewhere', every agent is reduced to a chatbot with good vocabulary.
And the 'somewhere' is the hard part. The invoice has to honor your tax setup. The stock movement has to respect FIFO and lot tracking. The payment has to land in double-entry without breaking the trial balance. Idempotency has to survive a retry. Rollback has to be auditable. Building this layer is months of careful work — and it is the part that decides whether your AI experiment becomes a real operation or stays a demo.
What Stripe Did for Payments, Handler Does for Operations
Stripe gave developers one battle-tested API for moving money. Handler gives agents the same kind of layer for everything else a business actually does — orders, stock, payroll, suppliers, customers, costing.
Before Stripe, every web app rebuilt its own payment plumbing — PCI, gateways, retries, refunds, disputes. Stripe collapsed that into a small set of well-modelled primitives behind a clean API, and an entire generation of products got to skip the hardest, most regulated part of their backend. Handler is doing the same thing, one layer up. The hard part for an AI-powered business is not the model. It is the operational layer the model has to act through.
Standard Business Entities
Order, Invoice, Payment, Stock Movement, Customer, Supplier, Product, Shift, Payroll Entry — fully modelled, with the relationships and statuses your bookkeeper expects.
Built-In Business Logic
createOrder() does the right thing — generates the invoice, reserves stock, updates the customer balance, posts the accounting entries. Agents call one function instead of five.
Audit Trail by Default
Every action is signed, timestamped and attributable to an agent and an approving human. Every posting is reversible through a documented rollback, never a silent overwrite.
Tool Registry, Not Just an API
Tools are exposed to agents through a typed function-calling registry — strict schemas, scoped permissions, idempotency keys built in. Hallucinated calls fail safely instead of corrupting your books.
The operational layer is what owns the truth. The model is a swappable piece on top. That separation is what lets the same business keep its operational integrity as the AI underneath gets faster, cheaper and smarter every quarter.
Agents Already at Work in Handler
These are not roadmap promises. These are agents that draft real documents in the system today, on real production data, with a human posting them in one tap.
Each agent inside Handler is small, scoped and inspectable. It has a defined job, a list of tools it is allowed to call, and a clear contract for what it produces. The agent never decides whether something happens — it decides what to draft. The decision to post still belongs to a human, and the audit trail records exactly who and when.
Ordering Agent
Watches days-of-supply, supplier lead times and seasonality. Drafts purchase orders to the right supplier, in the right quantity, at the right time. You approve.
Sales Agent
Reads B2B orders from LINE and WhatsApp. Resolves products, prices, customers. Drafts the sales order, the reservation and the invoice. You approve.
Payroll Agent
Pulls shifts, piece-rate, bonuses, deductions and anomalies. Drafts the period payroll with explanations per line. You approve and pay.
Receiving Agent
Turns supplier delivery photos and packing lists into stock receipt drafts with lot tracking and price reconciliation. You approve and post.
The architecture is open: new agents follow the same shape — scoped tools, drafted output, human posting, full audit. Adding a new agent does not require a new platform, just a new tool spec and a new prompt.
Every Agent Action Goes Through 5 Gates
There are no autonomous money decisions. The loop is designed so a human always owns the commit, and so every step is recoverable.
A draft is a real document inside Handler — same schema, same validation, same relationships as a posted one. The only difference is its status. It can be edited, rejected, re-drafted, or posted. Posting moves the document into the immutable history, generates the accounting entries, updates balances and inventory, and triggers downstream events. Nothing can be silently changed after that — only reversed through a documented rollback that itself becomes part of the audit trail.
Every step in the loop is recorded: what the trigger was, which tools the agent called with which arguments, which validations passed and failed, what the draft looked like, who approved it and when. If anything ever goes wrong, you can replay the exact decision the agent made, with the exact data it had at the time.
Agents Earn Their Permissions
Agents are not turned on with full power. They start as observers and earn the right to suggest, draft, then post — by track record, in scope.
| Level | Role | What the agent can do |
|---|---|---|
| L0 | Observer | Reads from Handler. Cannot change anything. Useful for monitoring and evaluation. |
| L1 | Suggester | Proposes actions in chat. Generates no documents and creates no drafts. Pure recommendation. |
| L2 | Drafter | Creates documents in draft state inside Handler. Always requires human posting before they become facts. |
| L3 | Co-Pilot | Auto-posts low-risk facts inside a scoped allowance (e.g. stock receipts ≤X, replenishment orders ≤Y). Larger actions still go through draft + approval. |
| L4 | Operator | Full posting authority inside its declared scope. Anomaly detection halts the agent automatically; every action keeps a 7-day human-rollback window. |
Promotions are explicit and reversible. Moving an agent from Suggester to Drafter is a one-click toggle by an admin. From Drafter to Co-Pilot requires you to set the scope (what kinds of facts can it post automatically) and the limit (how big a fact). Demotions are just as easy. The platform is designed so an over-eager agent in production cannot do more damage than its scope allows — and so an over-cautious one cannot block work that needs to happen.
What Changes for Your Business
Bringing the agent platform online is a phased shift, not a switch.
Visibility into the work agents can take
- See what agents would draft today, before they get any posting power
- Time the team currently spends on repetitive document work becomes measurable
- Quality of agent drafts becomes a number, not an opinion
- You decide which scopes you trust enough to move from Suggester to Drafter
Routine work moves from assembly to review
- Payroll and ordering become approvals, not multi-day spreadsheet sessions
- B2B order intake from LINE and WhatsApp stops bottlenecking on a single person
- Anomalies in price, supplier delays and scrap spikes are caught the same day, not next month
- The agent track record makes it safe to expand scopes and graduate the next agent
A business that documents its own knowledge
- Every approved agent action becomes part of the institutional memory
- New staff inherit clean workflows instead of inheriting tribal knowledge
- The owner stops being the bottleneck for routine decisions
- AI capability upgrades land as platform updates — no migration, no re-training your team
Most AI products give you a smarter chat. The agent platform gives you a smaller team for the same volume of work — without giving up control, audit or accounting accuracy. The data, the documents and the trail stay yours. The repetitive part stops being human work.
Your Data, Your Rules, Your Audit
Every action an agent takes is explainable down to the source data. You see the trigger, the tool call, the arguments, the validation results, the draft, the approving human and the resulting documents. There is no 'the model decided' that ends the explanation.
Agents can be paused per scope, per company, per minute. Permissions are tightened and expanded by configuration, not by code. The underlying LLM is a swappable piece — switching providers does not change the operational layer, the tool registry or the audit trail. Your moat is the data and the workflows on top of it, not whichever model runs this quarter.