Agents that reason, not workflows that execute.
Appstrate agents receive a goal and figure out the path. No graph to draw, no nodes to connect.
Workflows break the moment reality doesn't match the diagram.
Most AI platforms ask you to describe work as a graph. It works until it doesn't — a new edge case, a provider rename, a customer asking something slightly off-script.
Autonomous agents skip the graph. You define the goal and give the agent tools. On each run, it decides the steps. When the world changes, the agent adapts.
Goal in. Work out.
The agent reads the goal, plans execution on the fly, calls tools, and returns structured output.
{
"$schema": "https://afps.appstrate.dev/schema/v1/agent.schema.json",
"name": "@acme/refund-triage",
"version": "1.0.0",
"type": "agent",
"schemaVersion": "1.0",
"displayName": "Refund Triage",
"description": "Decide whether to refund, and how much.",
"dependencies": {
"providers": { "@appstrate/gmail": "^1.0.0" },
"tools": { "@acme/orders": "^1.0.0" }
},
"input": {
"schema": {
"type": "object",
"properties": { "email_id": { "type": "string" } },
"required": ["email_id"]
}
},
"output": {
"schema": {
"type": "object",
"properties": {
"decision": {
"type": "string",
"enum": ["approve", "reject", "escalate"]
},
"amount": { "type": "number" },
"reason": { "type": "string" }
},
"required": ["decision", "reason"]
}
},
"timeout": 180
}What makes it work.
Goal-driven
Agents receive high-level intent, not a step sequence. They plan on the fly.
Tool use
Pre-built tools (HTTP, filesystem, state, memory) + your custom TypeScript tools.
Skills
Reusable instructions packaged as SKILL.md files. Loaded per goal.
Memory & state
Agents persist state between runs. Multi-turn workflows without a vector DB.
Horizontal scaling without code
Each run is an isolated sandbox. No shared state, no coordination layer. Fan out to hundreds of concurrent runs — Appstrate handles the orchestration; you write zero queue code.
Ready to ship your first autonomous agent?
Self-host in 2 minutes. Or start from a template and adapt.