Early proof of concept · v0.1

Teach agents how to use software.

Agent Process is an open, portable format for describing how an AI agent navigates a web application and completes a task — step by step.

Not a runtime. Not another browser protocol. Just a simple way to describe a process.
Browser extension: download the latest build from the project folder and install it locally.

Agents shouldn’t have to rediscover how your software works on every run.

The idea

Tools tell agents what they can do. Processes tell them how to get work done.

Record it

Capture the real process in minutes

A person who already knows the application can simply perform the task once. A recorder can turn those clicks, fields, pages and expected states into an Agent Process.

Make it yours

Describe the process your team actually uses

The same software can be used in very different ways. Agent Processes can capture a company-specific path instead of relying on generic documentation or letting the agent figure everything out from scratch.

Less context

Give the agent the path, not the whole page

Instead of repeatedly inspecting large pages and reasoning over every available element, the agent can start with the relevant steps and targets. That can reduce unnecessary context and token usage while making execution more focused.

Portable

Record once. Export as structured data. Let compatible agents interpret the same process without depending on one vendor or model.

How it gets created

Record it once. Turn it into a process.

Anyone who knows the task can capture it directly inside the application with the Inline Manual Builder. No coding is required — the person recording it can be a developer, an implementation specialist, or simply the employee who performs the task every day.

1

Record

Open the Inline Manual Builder and perform the task normally. The Builder captures pages, steps, elements and interactions as you go.

2

Export

Turn the recorded flow into a portable Agent Process containing the relevant path, UI targets and expected states.

3

Agent follows it

Give the process to a compatible agent so it starts with the known path instead of figuring the application out from scratch.

Today: designed for browser agents and recorded web flows. Later: MCP, WebMCP, desktop, mobile and multi-application processes can build on the same model.
A tiny format first

Process → trigger → steps → elements → actions → expected result.

The first version is intentionally limited to web navigation. More complex execution models can come later without changing the core idea.

create_a_new_lead.agentprocess.jsonagentProcess: 0.1
{
  "description": "Creates a new lead in CRM",
  "domains": [
    "agentprocess.org"
  ],
  "enabled": true,
  "id": "a3fc22d2-cd90-4b4c-87ed-aa17311fab37",
  "name": "create_a_new_lead",
  "steps": [
    {
      "element": {
        "selector": "#newLeadBtn",
        "type": "button",
        "tagName": "BUTTON"
      },
      "title": "Click on New lead"
    },
    {
      "element": {
        "selector": "#firstName",
        "type": "text",
        "tagName": "INPUT"
      },
      "title": "Click on ",
      "variable": {
        "key": "first_name",
        "name": "First name",
        "type": "text"
      }
    },
    {
      "element": {
        "selector": "#lastName",
        "type": "text",
        "tagName": "INPUT"
      },
      "title": "Click on ",
      "variable": {
        "key": "last_name",
        "name": "Last name",
        "type": "text"
      }
    },
    {
      "element": {
        "selector": "#company",
        "type": "text",
        "tagName": "INPUT"
      },
      "title": "Click on ",
      "variable": {
        "key": "company",
        "name": "Company",
        "type": "text"
      }
    },
    {
      "element": {
        "selector": "#email",
        "type": "text",
        "tagName": "INPUT"
      },
      "title": "Click on ",
      "variable": {
        "key": "email",
        "name": "Email",
        "type": "text"
      }
    },
    {
      "element": {
        "selector": "#leadForm .primary",
        "type": "submit",
        "tagName": "BUTTON"
      },
      "title": "Click on Create lead"
    }
  ],
  "title": "Create a new lead",
  "version": "0.1"
}
A process in five parts

Simple enough to record from the UI.

01
Trigger

Where can this process begin?

02
Step

What should happen next?

03
Element

Which UI element is involved?

04
Action

Click, type, select, read, wait…

05
Expected result

How do we know the step worked?

v0.1 scope

Deliberately boring.

Web only.

Browser navigation and UI interaction first. Desktop, mobile, APIs and MCP bindings can come later.

Linear first.

Ordered steps before workflows, branching, loops and orchestration.

Semantic where possible.

Prefer roles, names, labels and stable identifiers. Keep CSS and XPath as fallbacks.

Recordable.

The format should map naturally to what browser recorders already capture.

Inspectable.

Plain JSON. Easy to diff, version, generate and validate.

Extensible.

Future versions can add branches, verification, WebMCP, MCP, approvals and multi-app processes.

Not trying to replace

MCP, WebMCP, browser automation or Agent Skills.

Agent Process is the description of the task. Execution can eventually be handled by whichever capability is available.

MCP

Expose tools and resources to agents.

WebMCP

Let a website expose structured actions directly to a browser agent.

Agent Process

Describe the sequence of work, UI context and expected outcome.

A possible future

A shared library of processes.

If the format proves useful, Agent Processes could become reusable building blocks. People, vendors and communities could publish processes for well-known software — or for specialised internal tools — and agents could discover and reuse them instead of learning every task from zero.

A company could still record its own version when its process differs from the default.

Create a Salesforce opportunity
Salesforce
Community process
Invite a user to HubSpot
HubSpot
Vendor process
Submit an expense report
Internal finance app
Private process
This is a proof of concept

We are starting with the smallest useful thing.

The goal is to test whether a simple, open representation of recorded web processes is useful to agent builders, browser automation projects and software vendors. The spec will change.