Skip to main content
A direct agent connects straight to the gateway. Start here unless you have a measured latency requirement that calls for sidecars.
Direct agents require a gateway: your own with Tilde OSS or Tilde Enterprise, or Tildeā€™s with Tilde Cloud.
A direct agent workload dialing out to the Tilde gateway, which forwards LLM requests to providers
Click to zoom
Your agent container holds your code and the Tilde SDK, in TypeScript or Python. The SDK dials out to the gateway over HTTP/2. Wakes, tool calls, and telemetry uploads all travel on that one connection. LLM requests also go to the gateway, which swaps in the real provider credentials and forwards them. The agent needs no inbound port, Kubernetes Service, or load balancer. It holds no provider API keys and needs no internet egress of its own.

ā€‹
Where direct agents run

A direct agent is an ordinary container, so it runs wherever you run containers:
  • Kubernetes, as a plain Deployment. It needs no Service or Ingress.
  • Managed container services, such as Amazon ECS and Fargate.
  • Any Docker host.
We recommend Kubernetes.

ā€‹
What a direct agent needs

Every replica of a release uses the same deployment token, so you scale by adding replicas. Replicas finish their active invocations on shutdown, so rolling updates do not interrupt conversations. Register each release from your pipeline. See Release agents from CI.