Vendor payment advice routing
Payment advices from the bank are read as they arrive, matched against the vendor master, and sent on to the right supplier.
AutoGo Agent Builder turns a plain-English description of a back-office task into a compiled, tested agent that runs as a service on your own servers. Every agent is built inside guardrails, tested on your sample files, and runs on Aurix, the same engine behind the AutoGo agents we deliver.
A model with open access to files, the network and credentials is hard to sign off. Nobody can say exactly what it will do on a bad day.
Recorded clicks are brittle, licences are per bot, and every fix waits on a specialist.
Every automation needs its own scaffolding for logging, retries, scheduling and deployment before any business logic gets written.
AutoGo Agent Builder takes a different approach. The AI writes the agent; the agent itself is ordinary compiled code with a fixed, audited set of things it is allowed to touch. You get the speed of describing work in plain English and the predictability of software your team can read, test and version.
| How the approaches compare | AI prototype agentmodel with open access | Screen RPArecorded clicks | Custom developmenthand-built service | AutoGo Agent Builder |
|---|---|---|---|---|
| Built from a plain-English description | ||||
| Access limits enforced, not just requested | ||||
| Keeps working when a screen layout changes | ||||
| Source code a developer can review | ||||
| Monitoring, retries and restart recovery built in |
The Builder handles the plumbing every agent needs (registration, heartbeat, job claiming, recovery after restart) so the generated code only contains your business logic.
Tell the Builder what should happen, in plain English, or start from a template. Point it at the reference files it needs, such as a vendor list or rate table.
The Builder writes the agent, compiles it, and reads its own build errors. Anything outside the guardrails fails the build and gets fixed.
Drop in real sample files. The agent runs in a sandbox with dummy secrets until the output is right. If something is unclear, the Builder asks you.
Export produces a single self-contained executable with encrypted settings and install instructions. No runtime to install.
The agent registers with the Aurix Control Centre and starts taking jobs from email, schedules, folders or the dashboard.
1[AgentMetadata(Id = "vendor-payment-router")] 2public class VendorPaymentRouter : AgentBase 3{ 4 public override async Task<JobOutput> ProcessAsync( 5 JobInput input, IAgentContext ctx, CancellationToken ct) 6 { 7 var advice = ParseAdvice(input.Files[0]); - var csv = File.ReadAllText(@"C:\finance\vendors.csv");+ var vendors = await LoadVendorsAsync(ctx.Data, ct);9 var vendor = vendors.Match(advice.Beneficiary); 10 ctx.Logger.Information("Matched {Vendor}", vendor.Name); 11 ...
Switch any session to Code view to read the agent's source, and use Diff to see exactly what the last message changed. Your developers review a single short file before anything reaches production.
ctx.Data (read-only)ctx.SecretsDrop in the files the agent will really receive, including the awkward ones. Test runs use sandboxed folders and secrets that exist only for the test.
ctx.Datactx.Browsersmtp.password via ctx.Secrets. Values are not exported. The README tells the operator where to set them.Pick Windows, Linux or macOS. Production folders and the Control Centre address are sealed into encrypted settings, and passwords never leave your hands.
Instructions to an AI can be ignored. A failed build cannot. Every agent can only reach the outside world through a small set of approved doors, and the build fails if the code tries any other route.
A compiler rule rejects direct file, network, process, registry and environment access, hardcoded paths and raw browser drivers.
Agents can use three vetted, version-pinned libraries for Excel, CSV and local OCR. Any other package fails the build.
Test runs are isolated child processes with a 5-minute timeout and memory and CPU caps, a file jail, and dummy secrets so real keys never reach a test.
Browser automation only reaches hosts you list. Server URL and agent token ship encrypted. Secret values are never exported with the agent.
Each agent is one C# file. Code view shows the source and a diff of every change, so developers can review it and keep it in source control.
| If generated code tries | The build fails and points to |
|---|---|
File.ReadAllText(@"C:\finance\vendors.csv") | ctx.Data.OpenReadAsync("vendors.csv") |
new HttpClient() | ctx.Http |
Environment.GetEnvironmentVariable("API_KEY") | ctx.Secrets.Get("erp.api_key") |
Console.WriteLine(...) | ctx.Logger |
| Agent | Status | Trigger | Last heartbeat | Jobs today |
|---|---|---|---|---|
| vendor-payment-router | Active | 12 s ago | 214 | |
| ar-journal-processing | Running | folder | 4 s ago | 532 |
| ar-matching | Active | schedule | 21 s ago | 389 |
| einvoice-submitter | Review queue | api | 9 s ago | 149 |
| sdr-outreach-drafter | Paused | manual | 2 h ago | 0 |
Route matching emails to an agent, run on a schedule, pick up files dropped in a folder, or submit by hand from the dashboard.
Each job records its inputs, progress, output files, metrics and final status. Failed jobs show the agent's own log lines.
Agents send a heartbeat, claim jobs from a central queue, and pick up where they left off after a reboot. That behaviour is part of the platform, not something each agent has to get right.
Illustrative dashboard. Figures are examples.
Agents call out to the Aurix Control Centre to claim work and report results. Nothing needs to call in to your servers.
ctx.FilesRead the files attached to a job and write output files back: PDFs, HTML emails, text and JSON.
ClosedXML · CsvHelperParse spreadsheets with odd header rows, reconcile ledgers, and produce formatted Excel reports.
TesseractPull text out of scanned documents on your own machine, without sending pages to an outside service.
ctx.Http · ctx.SecretsPost to ERP, e-invoicing and internal APIs, with credentials kept in a secrets store and out of the code.
ctx.BrowserHeadless browser automation for portals that have no API, restricted to hosts you allow and checked on every action.
ctx.Data · ctx.WorkspaceRead lookup tables like vendor lists or rate cards, and work inside a shared folder you nominate.
Payment advices from the bank are read as they arrive, matched against the vendor master, and sent on to the right supplier.
Incoming remittances are matched to open invoices and turned into journal entries ready for the ERP, with exceptions listed for a person to resolve.
Invoices are validated and submitted to the e-invoicing system, and every acknowledgement or rejection is kept with the job.
The agent researches a prospect and drafts a short, specific email. It only drafts: a person reads and sends every message.
Looking for a ready-made agent instead? See the AutoGo agents we deliver.
Agents install as a service on your servers or cloud VMs: Windows x64, Linux x64 and arm64, or macOS.
Agents you build register with Aurix just like the AutoGo agents we deliver: one queue, one dashboard, one job history.
Use the managed key that comes with your trial, or connect your own Anthropic account and keep billing in your name.
The Builder shows model usage and remaining trial credit while you work, so there are no surprise bills.
# What Export produces VendorPaymentRouter.exe ~70 MB, self-contained, no runtime to install appsettings.production.json encrypted: Control Centre URL, agent token, config README.txt install steps and the secret names this agent needs > VendorPaymentRouter.exe --install-service > sc start AIAgent.vendor-payment-router ✓ Registered with Control Centre · waiting for jobs
Aurix is the engine every AutoGo agent runs on. Agent Builder is how your team builds its own Aurix agents, instead of waiting for us to build each one. Agents you build sit alongside the ones we deliver, in the same Control Centre.
No. Analysts and operations leads build agents by chatting with the Builder and testing on sample files. Developers are useful for reviewing code before production, and the code is short and readable for that purpose.
Three things. The compiler rejects any code that touches files, the network, processes or the registry outside the approved context. Testing happens in a resource-capped sandbox with dummy secrets. In production, browser access is limited to hosts you list and credentials live in a secrets store, not in code.
While building, the task description and generated code are sent to the model. Your test files stay on the machine running the Builder. Deployed agents only call an AI model where a step needs one, such as summarising text, and only through the approved HTTP surface.
RPA replays recorded clicks and breaks when a screen changes. Aurix agents work with files, data and APIs directly, and use a browser only where a portal has no API. Each agent is one code file you can version and review.
Yes. Reopen the session, describe the change in chat, re-test on your samples, and export a new version. Your developers can also edit the code directly.
Agents always run on your infrastructure. The Aurix Control Centre is available as a managed service or installed in your own environment for data-residency requirements.
We'll build an agent with your team on one of your real back-office tasks, using your sample files, and show it running in the Aurix Control Centre.