Workflows

Workflows coordinate multiple agents into a sequential pipeline. Each step runs an agent, validates the output, and passes it to the next step via ctx.prev. Approval gates suspend the workflow until a human approves.

Defining workflows

Use the builder pattern: workflow() returns a builder, chain .step() calls, then .build() to finalize.

import { workflow } from '@vertz/agents';
import { s } from '@vertz/schema';

const pipeline = workflow('content-pipeline', {
  input: s.object({
    topic: s.string(),
    tone: s.string(),
  }),
})
  .step('research', {
    agent: researchAgent,
    input: (ctx) => `Research the topic: ${ctx.workflow.input.topic}`,
    output: s.object({ findings: s.string(), sources: s.array(s.string()) }),
  })
  .step('write', {
    agent: writerAgent,
    input: (ctx) => {
      // ctx.prev.research is typed as { findings: string; sources: string[] }
      return `Write about this using a ${ctx.workflow.input.tone} tone: ${ctx.prev.research.findings}`;
    },
    output: s.object({ draft: s.string(), wordCount: s.number() }),
  })
  .step('edit', {
    agent: editorAgent,
    input: (ctx) => `Edit this draft: ${ctx.prev.write.draft}`,
    output: s.object({ final: s.string() }),
  })
  .build();

Each step’s ctx.prev is strongly typed based on preceding steps’ output schemas. No casts needed.

Step options

Option	Type	Description
`agent`	`AgentDefinition`	The agent to execute. Omit for approval-only steps.
`input`	`(ctx: StepContext) => string \| { message }`	Transform workflow context into the agent’s message.
`output`	Schema	Schema for validating the agent’s response (parsed as JSON).
`approval`	`StepApprovalConfig`	Turns the step into a human approval gate.

Input callbacks

The input callback receives a StepContext with access to the workflow input and all previous step outputs:

.step('review', {
  agent: reviewerAgent,
  input: (ctx) => `Review this analysis: ${ctx.prev.analyze.summary}`,
  output: s.object({ approved: s.boolean(), feedback: s.string() }),
})

The callback can return a plain string or { message: string }:

// String form
input: (ctx) => `Analyze: ${ctx.workflow.input.topic}`,

// Object form
input: (ctx) => ({ message: `Analyze: ${ctx.workflow.input.topic}` }),

If no input callback is provided, the agent receives a default message: "Execute step \"step-name\"".

Workflow options

Option	Type	Description
`input`	Schema	Schema for validating workflow input.
`access`	`{ start?, approve? }`	Access control for starting or approving the workflow.

Validation rules

Workflow names must match /^[a-z][a-z0-9-]*$/
Step names must match the same pattern
At least one step is required
Duplicate step names within a workflow throw an error
Both workflow and step definitions are deeply frozen after creation

Running workflows

Use runWorkflow() to execute a workflow. It runs each step sequentially, passing outputs forward via ctx.prev.

import { runWorkflow, createAdapter } from '@vertz/agents';

const llm = createAdapter({ provider: 'openai' });

const result = await runWorkflow(pipeline, {
  input: { topic: 'TypeScript generics', tone: 'conversational' },
  llm,
});

if (result.status === 'complete') {
  console.log('All steps finished');
  console.log(result.stepResults);
} else if (result.status === 'error') {
  console.log(`Failed at step: ${result.failedStep}`);
  console.log(`Reason: ${result.errorReason}`);
  // 'agent-failed' — the agent hit max iterations or errored
  // 'invalid-json' — agent response wasn't valid JSON (but output schema expected it)
  // 'schema-mismatch' — agent response was valid JSON but didn't match the output schema
} else if (result.status === 'pending') {
  console.log(`Waiting for approval at: ${result.pendingStep}`);
  console.log(result.approvalMessage);
}

Result shape

type WorkflowErrorReason = 'agent-failed' | 'invalid-json' | 'schema-mismatch';

interface WorkflowResult {
  status: 'complete' | 'error' | 'pending';
  stepResults: Record<string, StepResult>;
  failedStep?: string; // Only set when status is 'error'
  errorReason?: WorkflowErrorReason; // Only set when status is 'error'
  pendingStep?: string; // Only set when status is 'pending'
  approvalMessage?: string; // Only set when status is 'pending'
}

interface StepResult {
  status: 'complete' | 'max-iterations' | 'stuck' | 'error';
  response: string;
  iterations: number;
}

How output accumulation works

Each step’s output is stored in ctx.prev keyed by step name. If a step has an output schema, the agent’s response is parsed as JSON and validated against it. Validated data is stored in prev. Steps without an output schema store { response: string }. If output validation fails (invalid JSON or schema mismatch), the workflow returns an error — it does not silently fall back.

Step 1 "research" completes → ctx.prev = { research: { findings: "...", sources: [...] } }
Step 2 "write" completes    → ctx.prev = { research: { ... }, write: { draft: "...", wordCount: 1200 } }
Step 3 "edit" completes     → ctx.prev = { research: { ... }, write: { ... }, edit: { final: "..." } }

ctx.prev is strongly typed — each step sees the output types of all preceding steps based on their output schemas. No casts needed.

Approval gates

An approval gate suspends the workflow until a human approves. When runWorkflow() hits an approval step, it returns immediately with status: 'pending'.

const reviewPipeline = workflow('review-pipeline', {
  input: s.object({ documentPath: s.string() }),
})
  .step('auto-review', {
    agent: reviewerAgent,
    input: (ctx) => `Review document at: ${ctx.workflow.input.documentPath}`,
    output: s.object({ approved: s.boolean(), findings: s.array(s.string()) }),
  })
  // Approval gate — no agent, just a gate
  .step('human-approval', {
    approval: {
      message: (ctx) =>
        `Auto-review found ${ctx.prev['auto-review'].findings.length} findings. Approve to proceed.`,
      timeout: '7d',
    },
  })
  .step('publish', {
    agent: publisherAgent,
    input: (ctx) => `Publish document at: ${ctx.workflow.input.documentPath}`,
    output: s.object({ url: s.string() }),
  })
  .build();

Approval config

Option	Type	Description
`message`	`string \| (ctx: StepContext) => string`	Message shown to the human approver.
`timeout`	`string`	How long to wait (e.g., `'7d'`).

Resuming after approval

When the workflow returns pending, store the step results. After the human approves, call runWorkflow() again with resumeAfter pointing to the approval step:

// First run — hits approval gate
const firstRun = await runWorkflow(reviewPipeline, {
  input: { documentPath: '/docs/api.md' },
  llm,
});

// firstRun.status === 'pending'
// firstRun.pendingStep === 'human-approval'
// firstRun.approvalMessage === 'Auto-review found 3 findings. Approve to proceed.'

// Store the results somewhere (DB, KV, etc.)
const savedResults = firstRun.stepResults;

// ... human approves ...

// Resume — skips all steps up to and including 'human-approval'
const resumed = await runWorkflow(reviewPipeline, {
  input: { documentPath: '/docs/api.md' },
  llm,
  resumeAfter: 'human-approval',
  previousResults: savedResults,
});

// resumed.status === 'complete' (if publish step succeeded)

The resumeAfter step name must match an existing step in the workflow. An invalid name throws an error.

Building the approval UX

The approval primitive is transport-agnostic — runWorkflow() doesn’t dictate how approvals are delivered or collected. Common patterns:

HTTP endpoint — Store pending state in a database, expose POST /workflows/:id/approve, render an approval button in a dashboard
Webhook — Send the approval message to Slack/Discord, listen for a reaction or command
Durable Object — On Cloudflare, hold workflow state in a Durable Object that wakes when an approval event arrives
CLI prompt — For dev tooling, prompt in the terminal and resume immediately

Agent-to-agent invocation

Tools can invoke other agents using ctx.agents.invoke(). This enables delegation patterns where a coordinator agent dispatches work to specialized agents.

const specialistAgent = agent('specialist', {
  state: s.object({}),
  initialState: {},
  tools: {
    /* specialist tools */
  },
  model: { provider: 'openai', model: 'gpt-4o' },
});

const delegateTool = tool({
  description: 'Delegate a task to a specialist agent',
  input: s.object({ task: s.string() }),
  output: s.object({ result: s.string() }),
  async handler(input, ctx) {
    const result = await ctx.agents.invoke(specialistAgent, {
      message: input.task,
    });
    return { result: result.response };
  },
});

const coordinator = agent('coordinator', {
  state: s.object({}),
  initialState: {},
  tools: { delegate: delegateTool },
  model: { provider: 'openai', model: 'gpt-4o' },
});

Invoke options

Option	Type	Description
`message`	`string`	The message to send to the target agent. Required
`instanceId`	`string`	Optional instance ID for the invoked agent.

The invoked agent runs a full ReAct loop with the same LLM adapter as the calling agent. It returns { response: string }.

Session persistence

Agents support persistent sessions via an AgentStore. Pass a store to run() to enable conversation history across multiple calls.

import { run, memoryStore } from '@vertz/agents';

const store = memoryStore();

// First message — creates a new session
const first = await run(greeter, {
  message: 'Hi, my name is Alice',
  llm,
  store,
});

console.log(first.sessionId); // 'sess_abc123...'

// Second message — resumes the session
const second = await run(greeter, {
  message: 'What was my name?',
  llm,
  store,
  sessionId: first.sessionId,
});
// Agent remembers the conversation

Available stores

Store	Import	Description
`memoryStore`	`@vertz/agents`	In-memory, for testing and dev.
`sqliteStore`	`@vertz/agents`	SQLite-backed via `vertz:sqlite`.
`d1Store`	`@vertz/agents`	Cloudflare D1 — pass `{ binding: env.DB }`. Tables are created on first call.

Session options

Option	Type	Description
`store`	`AgentStore`	The persistence backend. Required for sessions
`sessionId`	`string`	Resume an existing session. Omit to create new.
`maxStoredMessages`	`number`	Cap messages per session (default: 200).
`userId`	`string`	Session ownership — enforced on resume.
`tenantId`	`string`	Tenant scoping — enforced on resume.

LLM adapters

Agents communicate with LLMs through adapters. Use createAdapter() to create one:

import { createAdapter } from '@vertz/agents';

const llm = createAdapter({ provider: 'openai' });

Available providers

Provider	Value	Env variable
OpenAI	`'openai'`	`OPENAI_API_KEY`
Anthropic	`'anthropic'`	`ANTHROPIC_API_KEY`
Cloudflare AI	`'cloudflare'`	`CLOUDFLARE_ACCOUNT_ID`, `CLOUDFLARE_API_TOKEN`
MiniMax	`'minimax'`	`MINIMAX_API_KEY`

Custom adapters

You can provide a custom LLMAdapter directly — any object with a chat() method:

const customLlm: LLMAdapter = {
  async chat(messages, tools) {
    // Call your LLM and return the response
    return { text: '...', toolCalls: [] };
  },
};

const result = await run(myAgent, { message: 'Hello', llm: customLlm });

Agent lifecycle

Agents have three lifecycle hooks:

const myAgent = agent('monitored', {
  state: s.object({ startedAt: s.string() }),
  initialState: { startedAt: '' },
  tools: {
    /* ... */
  },
  model: { provider: 'openai', model: 'gpt-4o' },

  onStart(ctx) {
    ctx.state.startedAt = new Date().toISOString();
    console.log(`Agent ${ctx.agent.name} started`);
  },

  onComplete(ctx) {
    console.log(`Agent ${ctx.agent.name} completed`);
  },

  onStuck(ctx) {
    console.log(`Agent ${ctx.agent.name} got stuck`);
  },
});

Hook	Called when
`onStart`	Before the ReAct loop begins.
`onComplete`	After the loop completes successfully.
`onStuck`	When the agent hits `max-iterations` or `stuck` state.

Loop configuration

Control the ReAct loop behavior:

agent('careful', {
  // ...
  loop: {
    maxIterations: 50, // Max iterations before stopping (default: 20)
    onStuck: 'retry', // 'stop' | 'retry' | 'escalate' (default: 'stop')
    stuckThreshold: 3, // Iterations without progress before stuck (default: 3)
    checkpointInterval: 5, // Save state every N iterations (default: 5)
  },
});

Getting Started

@vertz/ui

@vertz/ui-compiler

@vertz/icons

@vertz/schema

@vertz/errors

@vertz/server

@vertz/agents

@vertz/db

@vertz/desktop

@vertz/fetch

@vertz/openapi

Deployment

Testing

Runtime

Defining workflows

Step options

Input callbacks

Workflow options

Validation rules

Running workflows

Result shape

How output accumulation works

Approval gates

Approval config

Resuming after approval

Building the approval UX

Agent-to-agent invocation

Invoke options

Session persistence

Available stores

Session options

LLM adapters

Available providers

Custom adapters

Agent lifecycle

Loop configuration

​Defining workflows

​Step options

​Input callbacks

​Workflow options

​Validation rules

​Running workflows

​Result shape

​How output accumulation works

​Approval gates

​Approval config

​Resuming after approval

​Building the approval UX

​Agent-to-agent invocation

​Invoke options

​Session persistence

​Available stores

​Session options

​LLM adapters

​Available providers

​Custom adapters

​Agent lifecycle

​Loop configuration

Defining workflows

Step options

Input callbacks

Workflow options

Validation rules

Running workflows

Result shape

How output accumulation works

Approval gates

Approval config

Resuming after approval

Building the approval UX

Agent-to-agent invocation

Invoke options

Session persistence

Available stores

Session options

LLM adapters

Available providers

Custom adapters

Agent lifecycle

Loop configuration