Workflows vs. agents

How to choose between deterministic workflows, AI agents, and the most powerful pattern: agents that call workflows as tools.

GTM Engine has two automation primitives:

Workflows are deterministic process graphs. They run the same steps every time for the same inputs.
Agents are AI assistants with a prompt, model, context, and tools. They decide which tools to call based on the conversation.

The important part: agents and workflows are not competing choices. The strongest automations usually use both. Put the repeatable business process in a workflow, then let an agent choose when to run it, explain it, or combine it with other tools.

One nuance: a workflow's graph and configuration are controlled, but AI Prompt and Agent tasks inside the graph can still produce model-variable output. Use schemas, validation, and downstream task settings when a result needs to be constrained.

User, trigger, or schedule

Something asks GTM Engine to do work: a record changes, a batch starts, or a user asks a question.

Agent decides

If the request is open-ended, an agent interprets intent and chooses the right tool.

Workflow executes

The workflow runs the precise steps: fetch records, enrich, score, write fields, send to an integration.

Result returns

The user gets an answer, the CRM is updated, or another workflow continues with structured outputs.

Workflows provide the reliable execution path. Agents provide the judgment layer that decides which path to use.

The Short Version

Workflow

Use a workflow when the path matters

Choose a workflow when the process should be repeatable, reviewable, testable, and versioned. If you can draw the steps as a flowchart, it probably belongs in a workflow.

Agent

Use an agent when the choice matters

Choose an agent when the user asks in natural language and you want the system to decide which action, search, report, or workflow tool to use.

Best pattern

Use both when you need judgment plus control

Build workflows for the actions you trust, then expose them to an agent as tools. The agent chooses; the workflow executes safely.

Rule of thumb

Avoid hiding business logic in prompts

If a rule must always be followed, put it in workflow configuration, field mappings, task settings, or permissions. Prompts are best for judgment, interpretation, and explanation.

Decision Guide

Use this path when you are designing a new automation:

1
Does this need to run automatically from an event, schedule, webhook, batch, or record action?
Triggers, manual runs, nested calls, and batch operations all invoke workflows cleanly.
Start with a workflow
2
Do the steps need to happen in a specific order with predictable inputs and outputs?
A workflow gives you a visible graph, typed variables, diagnostics, run history, and published versions.
Use a workflow
3
Will a user ask open-ended questions or expect the system to choose from several actions?
Agents are better at interpreting intent, choosing tools, asking follow-up questions, and explaining results.
Use an agent
4
Does the agent need to perform a real operation, not just answer?
Expose the operation as a workflow tool so the agent can call a controlled, versioned process.
Use both

At a Glance

	Workflow	Agent
What is it?	A directed graph of tasks (steps)	An LLM with a system prompt, model, and tool set
When does it run?	When a trigger fires, when called from another workflow or agent, or when a user runs it manually	When a user (or another agent) sends it a message, or when invoked as a tool by another agent
How is it configured?	Visual builder — drag tasks, wire variables, publish a version	Configure prompt, model, temperature, max steps, input variables, allowed tools
Determinism	Same inputs produce the same path through the graph	The model decides which tools to call and in what order
Best at	Repeatable processes you want to run the same way every time	Open-ended help, conversational interfaces, picking the right action from many options
Auditability	Strong: graph, published versions, task inputs/outputs, run history	Strong for conversation/tool traces, but less deterministic by design
Where it lives	`/automation/workflows`	`/automation/agents`

Use a Workflow When

Reach for a workflow when the problem sounds like "do this exact process every time."

Deterministic update

Auto-fill a CRM field

Read recent meetings, call an AI prompt, produce structured output, and write the result to a typed field.

Repeatable process

Enrich accounts or contacts

Call enrichment vendors, normalize the response, update records, and record errors step by step.

Batch-safe

Run a bulk operation

Apply the same published workflow to 200 accounts, rerun failures, and inspect run history later.

Pipeline

Score propensity

Fetch signals, run research, calculate scores, write reasoning, and keep the process consistent across records.

Controlled generation

Build outbound content

Generate email/social copy, validate outputs, write fields, and optionally enroll contacts in an integration.

Event-driven

React to a trigger

When a record changes or a schedule fires, map trigger outputs into workflow inputs and run the graph.

Workflows are best when you care about:

Control — every step is visible and versioned.
Validation — inputs, outputs, variable references, and schemas can be checked.
Run history — you can inspect what happened at each step.
Repeatability — the same process runs across records, batches, and triggers.

Use an Agent When

Reach for an agent when the problem sounds like "understand what the user means and choose the right action."

Conversational help

Page-aware Genie

A Forecast, Account, or Reports assistant can use page context plus tools to answer questions and suggest next actions.

Natural language interface

Slack or Teams assistant

A rep can ask for pipeline context, record lookup, or meeting prep without navigating the app first.

Tool choice

Research assistant

The agent can decide whether it needs web research, CRM records, saved reports, or a workflow based on the request.

Open-ended analysis

Ops copilot

The user can ask why pipeline moved, which deals need attention, or what action to take next.

Judgment plus action

Workflow launcher

The agent interprets intent, gathers missing inputs, and calls the right workflow tool.

Interactive

Guided form filler

The agent can ask clarifying questions and set fields based on the user's goal.

Agents are best when you care about:

Intent — the user may phrase the same goal many ways.
Choice — the system must pick among search, reports, workflow tools, and explanations.
Conversation — follow-up questions and clarifications are part of the experience.
Synthesis — the answer may combine product knowledge, CRM data, and research.

Use Both When

The common production pattern is agent as interface, workflow as action.

User asks

Run account research for Acme, then draft a follow-up plan.

Agent interprets

The agent identifies the account, chooses tools, and asks for any missing inputs.

Workflow tool runs

A published workflow performs the research, enrichment, scoring, and field writes.

Agent explains

The agent summarizes what happened and offers the next useful action.

This pattern keeps business operations controlled while still giving users a flexible natural-language experience.

Good examples:

A forecast Genie answers "which deals got riskier this week?" by searching records and reports, then calls a workflow to generate follow-up tasks.
A meeting-prep agent reads the account, recent activity, and web research, then runs a workflow that writes structured prep notes back to the CRM.
A Slack assistant lets a manager ask for "the top 10 accounts to prioritize today" and calls a propensity workflow to produce the ranked list.
A workflow can also call an Agent task when a repeatable process needs a judgment step in the middle, such as interpreting a transcript before choosing one of several branches.

Common Mistakes

Prefer workflow

Using an agent for a fixed ETL process

If every run should fetch, transform, validate, and write the same fields, use a workflow. The graph gives you diagnostics and run history.

Prefer agent

Using a workflow for open-ended Q&A

If the user might ask anything from record lookup to strategy questions, use an agent with a narrow set of useful tools.

Keep tools small

Giving an agent every tool

Agents work best with the smallest tool set that can complete the job. Too many broad tools make behavior harder to predict and test.

Encode rules

Putting required rules only in a prompt

Prompts guide behavior, but required business rules belong in workflow configuration, schemas, permissions, task settings, or validation.

Core Tools and Workflow Tools

Both Genie and custom agents draw from a catalog of tools. Tools usually fall into two buckets:

Core tools — built-in capabilities like product_help, search_records, search_reports, research_web, and navigate_to_record.
Workflow tools — workflows in your org with Use as Tool turned on.

When you add a workflow as a tool, it appears in the agent's tool list alongside core tools. That is the main bridge between controlled automation and conversational AI.

Where They Live

Workflows: /automation/workflows — list, builder, published versions, draft edits, folder organization, and the Use as Tool toggle.
Agents: /automation/agents — list, configuration page, allowed tools, and test chat.
Triggers: managed in the Automation Library; one trigger can fan out to many workflows, and one workflow can be invoked by many triggers.
Process Map: a canvas view of triggers, workflows, and agents wired together.

Next Steps

Read Workflows for the builder, versioning, triggers, inputs, outputs, and task graph.
Read Agents for prompts, models, tools, Genie, and test chat.
Follow Build your first workflow when you want a deterministic process.
Follow Build your first agent when you want a conversational assistant.
Read Use a workflow as an agent tool when you are ready to compose them.