What is an Agent?
Learn about What is an Agent? in WaymakerOS.
An Agent is a long-lived worker that remembers. Where an Ambassador wakes up, does a job, and forgets everything, an Agent keeps its memory between calls — the conversation so far, the workflow it's halfway through, the queue it's still draining.
Agents are the third thing you can build on Waymaker Host, alongside Apps and Ambassadors.
Addressed by id, not by web address
You don't visit an Agent. You address it by an id you choose — usually one per customer, per team, or per project:
waymaker agents invoke my-agent --id acme
waymaker agents invoke my-agent --id globex
Those two calls reach two completely separate instances of the same agent. Acme's instance has acme's memory. Globex's has globex's. Neither can see the other's data, and you don't write a line of code to keep them apart — the isolation is how Agents work.
This is what makes an Agent the right shape for per-customer work. One agent you build and deploy once, serving as many tenants as you have customers, each with its own private working memory.
What an Agent can do
| Capability | What it means |
|---|---|
| Durable memory | Store anything and it's still there on the next call — minutes or months later. Survives restarts and redeploys. |
| Self-scheduling | An Agent can set its own alarm: "wake me in an hour and follow up." No external scheduler needed. |
| AI built in | Generate, summarise, decide — using your Waymaker credits, or your own provider key (see below). |
| Your business data | Read and write your Commander tables, your app's database, and your organisation's context, scoped to your organisation. |
| Per-tenant isolation | Every id gets its own memory. Nothing leaks between tenants. |
When to build an Agent
Build an Agent when the thing needs to remember between interactions:
- A per-client assistant — one agent id per client, holding that relationship's history
- A workflow that waits — onboarding that spans days, chasing a step at a time
- A queue owner — something that accumulates work and drains it on its own schedule
- An ongoing conversation — where every message needs the ones before it
Build an Ambassador instead when the work is done the moment it finishes: a webhook handler, a nightly sync, a form processor. Build an App when people need a web address to visit.
The one-question test: does it need to remember? If yes, Agent.
Bring your own AI key
By default an Agent's AI runs through Waymaker and draws on your credits — nothing to configure. If you'd rather use your own provider account, set your provider and key on the agent and its AI calls go direct to that provider instead. You're then billed by your provider for the AI, and by Waymaker only for the time the agent spends running.
Your key is stored as a server-only secret: never visible to the agent's own code, never written to logs, and hidden whenever settings are read back.
What it costs
There's no limit on how many agents you create. Agents bill for what they use, not for existing:
- Running time — metered while the agent is actually working, in credits
- AI — metered per use if you're on Waymaker credits; billed by your provider if you bring your own key
- Memory — how much an agent can store depends on the licence of the person who owns it
Each agent has a billing owner — the person who created it, transferable to someone else — and consumption draws on that person's credit pool. If the pool runs dry, agents pause and their memory is kept safe; top up and they resume automatically. Nothing is ever deleted for non-payment.
Next Steps
- Creating Your First Agent — from scaffold to deployed
- Agent Memory and Schedules — state and self-scheduling
- Apps, Ambassadors, and Agents — which one do I build?