Talk to the work.
Don’t take our word for it — talk to our work. Every claim on this site eventually has to survive a conversation, and this page is where that happens: small, sandboxed agents built by the same practices we sell, doing real jobs in front of you. The roster below is in build right now; each agent goes live here the moment it passes the bar we’d hold it to at a client.
The roster
Five agents, five claims on trial.
-
The Reader
In buildPick one of three genuinely ugly documents — a crumpled invoice, a handwritten form, a bad scan — and watch it become clean fields, each carrying its confidence. One field gets held on purpose; that’s the point.
Proves Reading is content-based, and uncertainty routes to people.
IN BUILD — the sandbox opens with this agent.
-
The Concierge
In buildAsk anything about this company. Answers arrive with the page they came from, linked. Ask something we haven’t published and it says so — improvising isn’t in its permissions.
Proves Grounded retrieval, with “I don’t know” as a shipped feature.
IN BUILD — the sandbox opens with this agent.
-
The Router
In buildDescribe what hurts in your operation, in your own words. It names the practice that fits, shows its reasoning, and hands you the page — or tells you honestly that a consultation would route you better.
Proves Decision support with its work shown.
IN BUILD — the sandbox opens with this agent.
-
The Matcher
In buildThree documents: a purchase order, a packing slip, an invoice. One of them disagrees. The match runs, the hold opens with the difference marked — and you make the call. You’re the human in the loop.
Proves The three-way match, and why the hold exists.
IN BUILD — the sandbox opens with this agent.
-
The Spotter
In buildA month of transactions; one of them is lying. Set the sensitivity yourself and read why each flag was raised — in sentences, not a score.
Proves Anomaly detection you can interrogate.
IN BUILD — the sandbox opens with this agent.
Demos are promises with a user interface.
Which is why nothing here will ever be staged. The rules below are the build spec, not the fine print.
- Bounded
- each agent does exactly one job, deliberately.
- Grounded
- answers carry their sources or they don’t ship.
- Disposable
- nothing you type or pick is kept.
- Honest
- when it doesn’t know, it says so; we consider that a feature.
Questions
Asked about the sandbox.
Why publish agents that aren’t live yet?
Because the roster is a commitment with names on it. You can see exactly what we think is worth proving, and hold us to the list — which is roughly how our client work starts, too.
Will anything I type be stored?
No. The sandbox is disposable by design — no accounts, no transcripts kept, no training on your inputs. Bounded and forgettable is the spec.
Why these five?
One per claim: reading ugly documents, answering with sources, routing a symptom, holding a mismatch, explaining a flag. Each maps to a practice on this site — the roster is the services page, made conversational.
Could you build one of these for us?
That’s precisely the pitch. Every agent here is a scoped-down version of client work — the full-size versions are what AI Agents Development ships. Bring the job to a consultation.
What are they built with?
The same stack we sell: our practices, our guardrail harness, and DocuPOW where documents are involved. The sandbox isn’t a side project — it’s the showroom.
Will the roster grow?
When something new is worth proving, yes. Five honest agents beat fifteen demos — the bar stays where it is.