Execution, memory, tools, policies, evals and traces in one platform, so a prototype agent and a production one run on exactly the same rails.
Jobs survive restarts, retries and week-long waits. Every step is checkpointed, so nothing runs twice.
Vector and graph recall that every agent shares, fenced per customer so nothing leaks between tenants.
From Postgres to payroll, every tool call is typed in and typed out, with least-privilege credentials.
Rules live on the action, not in a prompt. A call that breaks one is denied before it runs.
Golden sets and graders run on every deploy, so a model change can never quietly break a workflow.
Traces, costs and decisions for every step, with one-click replay of exactly what each agent saw.
The pieces every team ends up building around an LLM, shipped as one runtime and maintained for you.
Checkpointed steps, retries and resumable waits without a queue to babysit.
Route each agent to the best model per step, with automatic fallbacks.
Every connector has a schema, so agents can't call a tool with the wrong shape.
Spend, scope and sign-off rules checked on every action, in four milliseconds.
Behaviour tests gate every release, the same way unit tests gate code.
Open any run and step through exactly what each agent saw and decided.
Every connector is typed in and typed out, with short-lived credentials scoped to the one agent that needs them.
Typed connectors
Describe the job, name the agents and the policies, set a budget. The runtime handles planning, retries, memory and the audit trail.
import { Ordin } from "@ordin/sdk"; const ordin = new Ordin({ region: "eu-west-1" }); const run = await ordin.run({ job: "Reconcile Q3 invoices", agents: ["planner", "analyst", "courier"], policies: ["spend-ceiling", "sign-off"], budget: { usd: 0.05 },}); console.log(run.status); // "settled"Free for the first 10,000 runs. Bring your own model, keep your own data, ship your first agent this afternoon.