Workflows vs Agents: Why Most AI Automation Is Over-Engineered

The Distinction That Matters

Here is the core framework:

Workflows are deterministic. They follow defined steps, handle structured inputs, and produce predictable outputs. Think: 'every Tuesday, pull new contacts from Apollo, format them, and add to the outreach queue.' The logic is fixed. The execution is repeatable. A cheap, fast model can do this.

Agents are autonomous. They receive a goal, decide how to achieve it, and exercise judgment along the way. Think: 'research this prospect and decide whether they are worth pursuing, then draft a personalized outreach angle.' The logic adapts. The output varies. This requires a more capable (and expensive) model.

The mistake most people make: they use agents for everything. They throw GPT-4 or Claude Opus at tasks that need a simple if-then pipeline. The result is slow, expensive, and unreliable -- because you are asking a thinking model to do a rules engine's job.

When to Use Workflows

Use workflows when:

The input is structured. CSV files, API responses, form submissions, database queries. If you know the shape of the data before it arrives, you want a workflow.

The logic is fixed. 'If the company has more than 50 employees and is in e-commerce, add to Sequence A.' This does not require judgment. It requires a conditional.

The output is predictable. Formatted emails, CRM updates, report generation from templates. If you could write a specification for every possible output, it is a workflow.

Speed and cost matter. Workflows run on cheap models (Haiku-class) or no model at all -- pure Python logic. They execute in milliseconds, not minutes.

Examples from my system: daily pipeline data pulls, contact formatting for Apollo enrollment, sitemap generation, content publishing (the git+deploy step), CRM data syncing.

When to Use Agents

Use agents when:

The input is ambiguous. A discovery call transcript, a competitor's website, a prospect's LinkedIn profile. The data is unstructured and requires interpretation.

The logic requires judgment. 'Is this prospect a good fit for our high-touch offering?' This is not a conditional -- it requires weighing multiple signals against a nuanced ICP definition.

The output should be creative or adaptive. Personalized outreach angles, strategic recommendations, content that needs voice calibration.

Error handling needs reasoning. When something goes wrong, the system needs to decide what to do next -- not just retry or fail.

Examples from my system: ICP research and segment identification (Alex), outbound copy quality gating (Bobbie), weekly strategic planning (Anna), revenue brief generation (Holden).

The Cost Difference Is Dramatic

This is not theoretical. Here are real numbers from my system:

A workflow that formats and enrolls contacts into Apollo sequences costs approximately $0.002 per execution. It runs on Haiku. It takes 3 seconds.

An agent that researches a prospect and drafts a personalized outreach angle costs approximately $0.15-0.30 per execution. It runs on Sonnet or Opus. It takes 30-90 seconds.

That is a 75-150x cost difference. If you are running the agent version on tasks that should be workflows, you are burning money. At scale -- hundreds of contacts per week -- this adds up fast.

The industry is moving toward what Anthropic calls 'model routing': using the cheapest model that can handle each specific task. Your architecture should reflect this. Not every task deserves your best model.

The Hybrid Architecture

The practical answer is not 'workflows OR agents.' It is both, with clear boundaries.

Here is how my system works:

Layer 1: Workflows (deterministic). Data collection, formatting, scheduling, publishing, CRM updates. These run on cron jobs (launchd on Mac Mini) with no AI model involved or with Haiku for light text processing.

Layer 2: Agents (autonomous, bounded). Research, analysis, drafting, quality review. These run in Claude Code sessions with defined roles, principles, and scope limits. They can make decisions but within guard rails.

Layer 3: Human (strategic, relational). Client calls, deal negotiations, editorial direction, final approvals on anything client-facing. This is where Gregor lives.

The routing rule is simple: start at Layer 1. Only escalate to Layer 2 when the task genuinely requires judgment. Only escalate to Layer 3 when the task requires relationship or strategic authority. Most tasks stay at Layer 1. That is the goal.

Why YC Is Funding AI-Native Agencies

This distinction explains a broader trend. Y Combinator is now funding AI-native agencies -- companies that deliver traditional services (marketing, legal, accounting) but with AI-first operations.

The reason: AI changes the economics of services businesses. When most of the execution is handled by workflows and agents, the margin structure starts to look like software (70-80%) instead of services (30-40%). You keep the revenue of a services business with the cost structure of a software company.

Adasight is already operating this way internally. The 10-agent team handles GTM execution at software-like marginal costs while delivering consulting services that command premium pricing. Productizing this for clients -- offering 'AI Workflow Sprints' as a service -- is the logical next step.

How to Audit Your Own Automation

Take every automated or AI-powered process in your business and ask three questions:

1. Is the input structured or unstructured? If structured, this should probably be a workflow.

2. Does the logic require judgment or just conditionals? If conditionals, this should definitely be a workflow.

3. What model is running this? If you are using Opus/GPT-4 for something that could run on Haiku or pure Python, you are over-engineering.

Most people will find that 60-70% of their 'AI agent' work should be workflows. Making this shift reduces costs, improves reliability (deterministic beats probabilistic for structured tasks), and reserves your expensive model budget for tasks that actually need reasoning.

The goal is not to use the most advanced AI everywhere. The goal is to use the right tool at each layer: proper routing = more automation at lower cost.