r.team Agents — how rAgent thinks
Loops and graphs, not prompts
Most AI tools work by typing a long instruction and hoping for a good answer. rAgent works differently: it looks things up in a map, checks its own work, and keeps going until the job is done right.
Why prompts run out of road
A prompt is a single instruction. You write everything the AI might need into one long message, press go, and get one answer back. That works for a poem. It fails for trust administration, for two reasons.
First, a prompt cannot try again. If the answer is slightly wrong, nothing catches it. Second, a prompt cannot hold structure. The rules of a trust, who the beneficiaries are, what was decided last year — all of that has to be pasted in as a wall of text every time, and the AI has to hope it read it correctly.
Loops solve the first problem. Graphs solve the second. The prompt does not disappear, but it shrinks to a few sentences describing rAgent's job and its limits.
What a loop is
A loop is the simplest idea in computing: do something, look at the result, decide whether it is good enough, and if not, do it again.
Instead of asking rAgent to be perfect in one try, we give it a way to check its own work and a clear rule for when it is finished. Good judgment comes from the checking, not from clever wording.
What a graph is
A graph is a map of things and the lines between them. A trust is connected to its trustee, its beneficiaries, its distribution rules, its assets, and every decision ever made about it.
When rAgent needs to know something, it does not reread a document. It follows the lines on the map and pulls only what matters for the question in front of it.
How rAgent puts them together
Take an everyday event: a beneficiary asks for money from a discretionary trust. Here is what happens, step by step.
- A request arrives The beneficiary's request lands in the Inbox. rAgent's job begins.
- rAgent reads the map It follows the graph from the beneficiary to the trust, to the standard that governs this kind of payment, to the assets available, to what was approved for this person before. Every fact it uses comes from a node it can point to.
- rAgent writes a proposal and checks it It drafts a recommendation and runs it against the firm's own rules and past decisions. If a check fails, it revises and goes back to the map. ↺ This is the loop. It repeats until the proposal passes.
- A person decides The proposal appears on the Approval screen. The trust company keeps every ounce of its discretion. rAgent recommends; it does not decide.
- The decision goes back on the map Once approved, the transaction is processed through the Bridge and the outcome is added to the graph as a new node. The next similar request will find it there.
Step 5 is the quiet one that matters most. Every approved decision teaches the map how this firm actually interprets its trusts. rAgent gets better without anyone rewriting a prompt.
What this means for a trust company
- Fewer made-up facts
- rAgent looks facts up rather than recalling them. If something is not on the map, it says so instead of guessing.
- A record you can replay
- Every proposal comes with the exact path rAgent walked through the graph and every pass around the loop. An examiner or a court can retrace it.
- Change a rule once
- When the firm updates how it reads a standard, the change is made on the map. Every future request sees it. Nothing needs re-prompting.
- A fair way to train and test
- In the Reasoning Agent School, an apprentice — human or rAgent — is scored on whether it followed the right lines on the map, not just whether it landed on the right number.
What still needs a person
Someone has to draw the map: decide which things and relationships belong on it and what they are called. Someone has to write the rules for when the loop is finished. And someone has to sit at the Approval screen. That work is harder than writing prompts, and it is the work r.team is built around.
Interactive demo — a phone walkthrough of propose → decide. See also Test in the lab, or About rAgent Team.