Real applications, built and running.

Credit agent: standard vs premium loan

A credit agent approves a standard loan when the credit score exceeds 650, and a premium loan when the customer has banked for more than 10 years. When both hold it generally grants the premium loan — unless current debt exceeds 60% of annual income, in which case it falls back to the standard loan. A textbook case of competing rules resolved by preference and exception.

Chest-pain triage from clinical data

Built directly from a tabular dataset, this agent decides a patient's disposition — discharge, observe, or admit — from the HEART score, troponin result and known coronary disease. Rules and preferences settle the competing signals, and an age-based meta-preference escalates the disposition for older patients. Every verdict carries its rationale.

Call assistant: how to handle an incoming call

A personal call assistant decides how to handle each incoming call from the user's context — at work, at home, in a meeting — the caller's relationship, and live signals. Assumptions (abducibles) stand in for what can't be observed, and the agent returns a decision with the reasoning behind it — according to the caller's feedback when needed.

The scenario I want my call assistant agent to choose one of two options: let the phone ring, or deny the call. Generally, I prefer the first option, unless the caller is unknown, in which case I always deny. At home, I accept calls from family and friends, but not from colleagues or my manager. However, I accept the call from my manager assuming there is an urgency that has to be confirmed by the caller. Similarly, if I am in a business discussion on the phone at home, I deny the call and pass on the message: "I'm in a business discussion, please call back later." At work, I accept calls from colleagues, friends, and family. However, if I am in a meeting, I prefer to deny the call and pass on the message: "I'm in a meeting, please call back later," unless it's a family member. In that case, I will accept the call assuming there is a health problem and the health issue is confirmed by the caller.

What to do tonight: deciding under missing information

A small everyday decision — stay home, go to a restaurant, or to the cinema — used to make visible how the engine treats the three kinds of input it reasons over: facts that hold (it's raining, I feel tired), beliefs derived under the asserted scenario when sources disagree (one forecast says rain, another says sunny — the engine resolves it based on the fact that the user considers BBC more accurate than France 2), and abducibles (assumptions) whose truth value matters but isn't yet established. When the verdict depends on an unconfirmed abducible, the engine doesn't guess: it surfaces the abducible as a yes/no question and re-decides once you answer.

The scenario When I believe that it will be a sunny day, I (prefer to) go to the restaurant. But when I feel tired, I (prefer to) stay at home, assuming that there is a nice movie on TV. Usually, if both situations occur, I prefer to stay at home. However, if it is my friend's birthday, I prefer to go to the restaurant with him. Today, France 2 weather forecast announced that it will be a sunny day, but BBC weather forecast announced that it will rain. I believe it's going to rain because I consider the BBC weather forecast to be usually accurate.

Trip advisor: recommendations you can question

The travel-data layer gathers candidate flights and hotels; the symbolic AI reasoner evaluates each under the traveller's profile and constraints. The result is a recommendation with a full argument trace — plus the abductive what-if alternatives. Each user layers their own personalisations on the fly, on top of the agency's generic policy without touching it — creating a personalized service for themselves while the shared behaviour stays the same for everyone.

In this run The agent automatically detects high temperatures forecast in Athens on the travel date, applies the user's personal policy — prefer evening flights when the destination is hot — and switches from the initially proposed morning departure to a late-evening one.

Note: the generated PAF code shown in this demo appears only to make visible that code is automatically generated on the fly from the user's policy. In a real-world deployment it is hidden.

Contract Sentinel: clause-by-clause review

Paste a contract and the agent extracts the relevant clauses, reasons over each against a curated argumentation playbook with risk-tolerance preferences, and returns a verdict — accept, negotiate, or reject — together with the verbatim passage that triggered it and an abductive what-if identifying the abducibles (assumptions) that, if their value changed, would flip the verdict.

Hiring agent: automated initial screen

From a written hiring policy, the agent screens a candidate and returns one of three suggestions — progress, decline, or hold for recruiter review — applying hard requirements, exceptions and counter-exceptions. Protected attributes are excluded by design, every verdict carries its audit trace, and the screen supports the recruiter without binding them.

KYC + SAR: two agents with complementary expertise

A multi-agent compliance system on a single customer file. A KYC agent decides whether to approve, decline or escalate the onboarding; a SAR agent decides whether to file a Suspicious Activity Report. Each has its own policy, vocabulary, and verdict set — complementary expertise rather than overlap. When the KYC verdict turns on something the SAR agent is better positioned to assess, the KYC agent surfaces an abducible (assumption) — "if SAR confirms X, my verdict becomes Y" — and asks the compliance officer to invoke SAR. The SAR agent runs on the same customer file, its verdict feeds back as a confirmed input, and KYC re-decides with the new information. The collaboration is short, explicit, and fully traceable.

Firewall agent: permit or block outbound traffic

A network firewall agent decides whether to permit or block outbound traffic to an external server. Its default is to permit — but a destination on the organisation's blocklist is blocked, and an application on the whitelist overrides that block and is permitted regardless. A compact policy of rule, exception, and counter-exception, resolved by preference, with every verdict carrying the trace that produced it.

The scenario A firewall agent can permit or block outbound traffic to an external server. Generally it permits. However if the destination server is on the organisation's blocklist it blocks the traffic. But if the traffic is initiated by an application on the organisation's whitelist it permits it regardless of the blocklist status.

Logistics agent: choosing a shipping method

Built directly from a tabular dataset, this agent chooses a shipment's method — standard ground, expedited air, economy sea, or hold for review — from its weight, distance, product importance, perishability, customs risk and cost sensitivity. In this run a perishable but very heavy, very short-haul shipment routes to economy sea. Three rules fire at once: the default standard-ground rule, an expedited-air rule (the item is perishable), and an economy-sea rule (it is very heavy on a very short route). Each pair conflicts, and three preferences settle them — expedited beats standard, economy beats standard, and, decisively, economy beats expedited when the shipment is very heavy and the distance very short. That last preference flips what would normally fly onto a ship. The full argument trace shows every rule that fired, the conflicts between them, and the preference that decided the outcome.

The scenario Generally a shipment goes by standard ground. A high-importance item travelling a long distance is sent by expedited air. A heavy, cost-sensitive shipment goes by economy sea instead. When an item is perishable on a short route, the agent prefers economy sea to protect it. But whenever the customer's file is flagged for customs risk, the shipment is held for review — this takes priority over every other choice.