Agents inside software

An agent that knows your software.

An agent added from the outside must infer rules hidden in screens and code. Built into the software, it has explicit actions, its user’s permissions and working context.

Explore

The screen does not show every rule.

An organization may forbid a region. A field may depend on a record’s status. An operation may require approval. Reading HTML does not necessarily reveal any of this.

  • Business rules and allowed values
  • The person’s identity and the agent’s identity
  • History, previous state and undo
  • The boundary between received data and an authorized instruction

Every screen has an agent contract.

A person chooses a value in an interface. An agent calls a named action exposing the field, allowed values and required permission. Both go through the same business logic.

When the design changes, the action contract can stay stable. The agent does not have to relearn where a button lives.

Chat is one entry point among others.

A signed contract can trigger a summary and a letter. A release can prepare its documentation. A refund request can be evaluated under company rules, then handed to a person when needed.

The software decides when the agent gets involved. The result returns to the record, process or editor where the team already works.

01 — In action

Your software starts it.
Syrtis prepares the result.

Change the data, run the agent and see how it prepares the work. You stay in charge of approval.

Simulation · fictional data

From signature to a ready record.

Change the data, then run the scenario.
01

Your trigger

Above €25,000 total commitment, a manager must review.

02

The agent orchestrates

Data

The result of this step will appear after it runs.

03

Your result

Ready when you are.

The draft will appear here. You can edit it before approving.

Action history0
  1. The history fills as the process runs. It includes each approval and undo action.

03 — Interfaces

The same software.
Two interfaces.

Julie sees a map and fields. The agent receives their actions, values and rules. They work on the same record.

Julie → her agentSimulation · fictional data

« Prepare the Dupont renewal. »

Drag to discover what the agent understands.

Same field. Same rules. Two representations.

The contract the agent reads

Dupont SARL · explicit actions and rules

ClientRecord.agent.yamlYAML
Fields and values
component: "ClientRecord"on_behalf_of: "Julie"approval: "required_before_save" input:  timezone:    value: "Europe/Brussels"    format: "IANA"   discount:    value: 0    format: "number"   reminder:    value: ""    format: "YYYY-MM-DD"   note:    value: ""    format: "string" 
Actions and constraints
tools:  pick_timezone:    target: "input.timezone"    params: { zone: IANA }    allowed:      - "Europe/Brussels"      - "Europe/Paris"      - "America/New_York"      - "Asia/Tokyo"   contract.set_discount:    target: "input.discount"    params: { percent: number }    range: [0, 20] # Julie’s limit   reminder.schedule:    target: "input.reminder"    params: { date: YYYY-MM-DD }   note.write:    target: "input.note"    params: { text: string }    max_length: 500 save:  approval: "required_before_save"
component: "ClientRecord"on_behalf_of: "Julie"approval: "required_before_save" input:  timezone:    value: "Europe/Brussels"    format: "IANA" tools:  pick_timezone:    target: "input.timezone"    params: { zone: IANA }    allowed:      - "Europe/Brussels"      - "Europe/Paris"      - "America/New_York"      - "Asia/Tokyo" 

Customer record — Dupont SARL

Julie · Lyon branch

Choose a city. The hatched zone is excluded by the rule.

Julie’s limit: 20%

Renewal due: January 1, 2027

The agent proposes. Julie edits and approves.

Nothing is saved without Julie’s approval.
View the full JSON contract
{
  "component": "ClientRecord",
  "on_behalf_of": "Julie",
  "approval": "required_before_save",
  "fields": [
    {
      "field": "timezone",
      "value": "Europe/Brussels",
      "action": "pick_timezone",
      "params": {
        "zone": "IANA"
      },
      "format": "IANA",
      "rule": "Brussels, Paris, New York or Tokyo. Other zones are excluded in this example.",
      "allowed": [
        "Europe/Brussels",
        "Europe/Paris",
        "America/New_York",
        "Asia/Tokyo"
      ]
    },
    {
      "field": "discount",
      "value": 0,
      "action": "contract.set_discount",
      "params": {
        "percent": "number"
      },
      "format": "Percentage",
      "rule": "From 0 to 20%. The agent cannot exceed Julie’s limit.",
      "min": 0,
      "max": 20
    },
    {
      "field": "reminder",
      "value": "",
      "action": "reminder.schedule",
      "params": {
        "date": "YYYY-MM-DD"
      },
      "format": "YYYY-MM-DD",
      "rule": "An explicit date, shown to Julie before saving."
    },
    {
      "field": "note",
      "value": "",
      "action": "note.write",
      "params": {
        "text": "string"
      },
      "format": "Text · up to 500 characters",
      "rule": "A note Julie can edit before approval.",
      "max_length": 500
    }
  ]
}

It starts with
one process.

The one your team does by hand, every week.
What if your software took care of it?