Ch10. Subagents, Workflows, Remote Agents

Subagents are how Eve breaks work into smaller durable runs. Use them for quality, specialization, and authority reduction.

Two Subagent Types

Built-in agent Tool

The built-in agent tool launches a fresh copy of the current agent for parallel or focused work. It is useful for decomposition but is not a security boundary because it shares the parent's capability surface.

Use it for:

• parallel summarization
• drafting alternatives
• independent analysis of the same evidence

Avoid it when authority must be reduced.

Declared Subagent

agent/subagents/researcher/
├── agent.ts
├── instructions.md
├── tools/
└── skills/

import { defineAgent } from "eve";

export default defineAgent({
  description: "Research internal documentation and return cited findings.",
  model: "anthropic/claude-opus-4.8",
});

The description is required because the parent model uses it to decide when to delegate.

Isolation Boundary

This is why declared subagents are useful for least-privilege architecture. Give a researcher read-only connections and an operator approval-gated write tools.

What The Parent Sees

The parent receives the subagent result, not all internal reasoning. Use outputSchema when the parent needs a stable machine-readable contract.

outputSchema: z.object({
  finding: z.string(),
  citations: z.array(z.string()),
  risk: z.enum(["low", "medium", "high"]),
});

Remote Agents

Remote agents let one Eve deployment call another deployment as a subagent.

import { defineRemoteAgent } from "eve/agents";
import { vercelOidc } from "eve/agents/auth";

export default defineRemoteAgent({
  url: "https://data-reviewer.example.com",
  description: "Reviews warehouse analysis plans for privacy and cost risk.",
  auth: vercelOidc(),
});

Use remote agents when:

• ownership belongs to another team
• the capability has its own deploy cadence
• data boundaries require a separate service
• scaling characteristics differ

Remote Auth

Remote agents are production APIs. Treat them like external services with SLOs, auth, versioning, and incident ownership.

Experimental Workflow

ExperimentalWorkflow lets the model write JavaScript to orchestrate subagents inside a durable step.

export { ExperimentalWorkflow as default } from "eve/tools";

Use it narrowly:

• only subagent orchestration
• no direct filesystem/network/arbitrary imports
• no sensitive side effects inside workflow code
• strong eval coverage

Orchestration Patterns

Evaluation Criteria