are workflow step invocations mandatory?

[GabrielLunesu]

for a chat-style session that is already warm, is ~4 sequential Workflow step invocations before the first model call expected, and is there a supported way to reduce the pre-model step count or keep the sandbox warm between turns?

[PrithwishDas101]

The workflow steps are part of eve's durability model, but the number of steps before the first model call isn't necessarily fixed at four. It depends on the execution path and the work performed before the model is invoked.

A few things worth separating:

• Workflow checkpoint overhead: eve checkpoints progress at Workflow step boundaries. The documented experimental option experimental.workflow.modelCallsPerStep lets you batch sequential model calls and their inline tool calls into a single step. For example:

  import { defineAgent } from "eve";
  
  export default defineAgent({
    model: "anthropic/claude-sonnet-4.5",
    experimental: {
      workflow: {
        modelCallsPerStep: 4,
      },
    },
  });

  This is a trade-off rather than a free optimization: if a step is interrupted, earlier model calls and inline tool executions in that batch can be replayed. Keep non-idempotent side effects protected with idempotency keys.

• Sandbox warmness: the sandbox and the durable workflow have separate lifecycles. eve can park the workflow without holding sandbox compute, and the runtime opens or reuses sandbox compute when code needs it. Keeping the workflow session alive therefore doesn't, by itself, guarantee that sandbox compute stays warm.

• What to measure: instrument the workflow and sandbox boundaries separately to determine whether the delay comes from workflow dispatch/checkpointing, sandbox acquisition, or model-provider latency. Without a trace showing those timings, it's hard to identify which steps can safely be removed.

The relevant docs are [Execution Model and Durability](https://github.com/vercel/eve/blob/main/docs/concepts/execution-model-and-durability.mdx) and [Agent Configuration — Workflow checkpoint batching](https://github.com/vercel/eve/blob/main/docs/agent-config.md).

I'd start by measuring the current path, then test modelCallsPerStep if the workload has sequential model/tool cycles. I wouldn't assume there's a supported way to pin sandbox compute warm across turns without confirming that against the current sandbox backend.