Systems

The Agent Mental Model: The System Is the Product

An agent is not a chatbot and not a permanent employee. It's a temporary worker that wakes up, reasons, uses tools and stops. Here's the mental model I use to design every one of them.

By Mohanad Anan 4 min read

A bronze automaton sits at a long workshop table in a vaulted hall while a robed master, standing slightly above it, briefs it by pointing at an open ledger, with chisels, scrolls and a quill ready within its reach.
In this essay · 13 sections
  1. 01 What an Agent Really Is
  2. 02 The Core Formula
  3. 03 The Most Important Insight
  4. 04 How To Think About Every Agent
  5. 05 The Golden Rule
  6. 06 Why Agent SDK Is Powerful
  7. 07 Insight: Your system setup matters more than the model
  8. 08 The correct mental model
  9. 09 The Biggest Limitation
  10. 10 Recruiting Example
  11. 11 Accounting Example
  12. 12 Property Management Example
  13. 13 The Ultimate Mental Model

What an Agent Really Is

An agent is not an employee that lives forever.

An agent is a temporary worker that:

  1. Wakes up when triggered
  2. Receives a briefing
  3. Thinks
  4. Uses tools
  5. Produces actions
  6. Writes results back to memory
  7. Stops

The database remembers.

The agent reasons.

The Core Formula

Agent = Intelligence + Context + Tools + Rules

Without intelligence, it cannot reason.

Without context, it cannot understand the situation.

Without tools, it cannot affect the world.

Without rules, it becomes unpredictable.

The Most Important Insight

The model is not the product.

The system is the product.

Most people focus on:

  • GPT
  • Claude
  • Gemini
  • DeepSeek

The real leverage comes from:

  • Database design
  • Workflow design
  • Tool design
  • Memory design
  • Escalation design
  • Logging design

The model is only one component.

How To Think About Every Agent

Whenever building an agent, ask:

1. What wakes it up?

Examples:

  • New lead entered CRM
  • Customer sent message
  • Invoice became overdue
  • Daily 8:00 scan
  • Support ticket opened

2. What information does it need?

Examples:

  • Lead profile
  • Conversation history
  • Current stage
  • Customer preferences
  • Previous actions

Only load what matters.

3. What tools can it use?

Examples:

  • Send WhatsApp
  • Send email
  • Book appointment
  • Update database
  • Create invoice
  • Search knowledge base

4. What decisions may it make?

Examples:

  • Follow up
  • Escalate
  • Book call
  • Send resource
  • Mark qualified

5. What must never happen?

Examples:

  • Promise something false
  • Spam users
  • Delete records
  • Spend money without approval
  • Ignore opt-outs

The Golden Rule

Use code for certainty.

Use agents for judgment.

Bad:

“Agent decides everything.”

Good:

“Code handles rules. Agent handles reasoning.”

A split composition: a precise brass clockwork mechanism turning on the left, a contemplative sage weighing a decision with a raised hand on the right, joined by a bridge of light.
Certainty belongs to code, judgment to the agent, and the design lives in the bridge between them.

Why Agent SDK Is Powerful

Traditional software:

Input → Logic → Output

Agent software:

Input → Reasoning → Tools → Decisions → Output

Instead of programming every possible path, you define:

  • Goal
  • Rules
  • Tools
  • Permissions

Then the agent figures out many situations dynamically.

This allows software to handle cases that were never explicitly programmed.

Insight: Your system setup matters more than the model

This is the big insight.

The agent is only as good as the environment you give it.

You need:

Clear rules
Clear tools
Clear permissions
Clear lead stages
Clear escalation logic
Clear database schema
Clear logging
Clear memory summaries
Clear workflows

Then the reasoning model becomes powerful.

Without that, the agent becomes a smart person dropped into chaos.

The correct mental model

The model = intelligence
The Agent SDK = execution framework
The tools = hands
The database = memory
The rules = law
The workflow = map
The logs = accountability
The human = final authority

That’s the architecture.

An allegorical fresco of a mechanical figure holding craftsman's tools, its head softly lit, with a great archive book at its feet, stone tablets of law behind it, an unrolled map before it and a robed human enthroned above as judge.
Each part has one job: the model thinks, the tools act, the database remembers, and a person has the final word.

The Biggest Limitation

Agents do not truly understand your business.

They only understand:

  • The context you provide
  • The tools you expose
  • The rules you define

Poor context produces poor decisions.

Poor workflows produce poor outcomes.

Most failures are system failures, not model failures.

Recruiting Example

Goal:

Convert interested candidates into booked interviews.

Trigger:

Candidate sends WhatsApp message.

Context:

  • Candidate profile
  • Job information
  • Recent messages
  • Current stage

Tools:

  • WhatsApp
  • Calendar booking
  • CRM updates

Possible actions:

  • Answer question
  • Handle objection
  • Book interview
  • Escalate to recruiter

Result:

One recruiter can manage dramatically more candidates.

Accounting Example

Goal:

Collect unpaid invoices.

Trigger:

Invoice becomes 14 days overdue.

Context:

  • Customer information
  • Invoice amount
  • Payment history

Tools:

  • Email
  • CRM
  • Accounting software

Possible actions:

  • Send reminder
  • Escalate
  • Offer payment plan
  • Notify finance team

Result:

Most collections happen automatically.

Property Management Example

Goal:

Handle tenant issues.

Trigger:

Tenant submits maintenance request.

Context:

  • Tenant
  • Property
  • Maintenance history
  • Lease information

Tools:

  • Email
  • SMS
  • Contractor database
  • Ticket system

Possible actions:

  • Classify issue
  • Contact contractor
  • Schedule visit
  • Update tenant

Result:

Property managers spend time on exceptions instead of administration.

The Ultimate Mental Model

An agent is not a chatbot.

An agent is a reasoning layer sitting between:

Events → Memory → Tools → Actions

The database stores reality.

The agent interprets reality.

The tools change reality.

The workflow controls reality.

Mohanad

Düsseldorf, 19 June 2026