AI agents

Runlane can coordinate an AI support agent inside a TypeScript task. The agent can search product documentation, request approval before a refund, and return a typed response. Your application supplies the model, tools, credentials, approval policy, and conversation state.

Run one durable support turn

The Runlane calls below are real. The marked runSupportAgent() call is pseudocode for your model and tool implementation:

src/tasks/support-agent.ts

import { task } from '@runlane/core'
import * as z from 'zod'

const messageSchema = z.object({
  content: z.string(),
  role: z.enum(['assistant', 'user']),
})
const supportResultSchema = z.object({
  reply: z.string(),
  toolCalls: z.array(z.enum(['refundOrder', 'searchDocs'])),
})

export const supportAgent = task({
  id: 'agent.support',
  schema: z.object({ messages: z.array(messageSchema).min(1).max(50) }),
  output: supportResultSchema,
  retry: { maxAttempts: 3 },
  async run({ messages }, context) {
    return context.step.run('support-turn', { output: supportResultSchema }, async ({ token }) => {
      //////////////////////////////////
      // MAKE YOUR API CALLS HERE
      //////////////////////////////////
      const result = await runSupportAgent({
        approvalRequiredFor: ['refundOrder'],
        idempotencyKey: token,
        messages,
        signal: context.signal,
        tools: ['refundOrder', 'searchDocs'],
      })

      return supportResultSchema.parse(result)
    })
  },
})

The durable step stores the completed support turn. If a later attempt reaches the same step, Runlane returns the stored result instead of intentionally running the agent again. Pass the step token to providers that support idempotency.

approvalRequiredFor describes a tool policy owned by your agent integration; it is not a Runlane API. Use a wait token when approval should end the current attempt and resume the run later.

Choose the right task shape

Pattern	Runlane approach	Limit
Support turn	Save one bounded turn as a durable step	The step replays as one unit
Prompt chain	Use one validated step per stage	Changing a step key creates new work
Parallel calls	Use `Promise.all()` inside one attempt	It is not a durable child-run join
Tool loop	Use stable step keys for costly calls	Cap turns, cost, and tool calls
Human review	Wait on a durable token	Your app owns authentication and the review UI

Runlane does not include an agent SDK, prompt store, chat session, or live frontend stream. A long in-memory loop holds a worker slot.

Before production

Put model rate limits on a bounded queue.
Keep secrets and full conversations outside run metadata.
Validate every saved model or tool result.
Require authorization before tools perform sensitive actions.
Test a retry after a completed step and confirm the provider is not called again.

Read save step results, limit concurrency, and human in the loop.

Run one durable support turn

Choose the right task shape

Before production

On this page