Agent Engineering Patterns

Patterns are reusable solutions to recurring agent design problems.

Use this catalog when you know the symptom but are not sure which architecture to choose.

Pattern Index

Pattern	Use When	Avoid When
Single-Purpose Agent	One narrow task is enough	The task requires external actions
Tool-Using Agent	The model needs external capabilities	The tool has high-risk side effects and no approval gate
RAG Agent	The answer depends on external knowledge	Retrieval quality cannot be evaluated
Memory Agent	Stable context should persist across turns	The information is sensitive or unstable
Planner-Executor	The task has multiple steps	The task is simple enough for one call
Reviewer-Evaluator	Quality must be checked before delivery	The rubric is vague
Router-Specialist	Different task types need different handlers	Routing is ambiguous and untested
Supervisor Multi-Agent	Several specialists must collaborate	No one owns the final answer
Human Approval Gate	Actions are high-impact or irreversible	The action is low-risk and frequent
Evaluation Harness	You need regression protection	You have no expected behavior defined

1. Single-Purpose Agent

Problem

You want an agent that reliably performs one task.

Structure

Input -> Prompt + Rules -> Model -> Structured Output -> Validation

Use For

• summarization
• classification
• extraction
• rewriting
• planning drafts

Failure Mode

The scope gets too broad and the agent starts pretending to be a general assistant.

2. Tool-Using Agent

Problem

The model cannot complete the task with text alone.

Structure

Input -> Decide Tool -> Validate Arguments -> Execute Tool -> Interpret Result -> Answer

Use For

• calculation
• search
• database lookup
• document operations
• external API calls

Failure Mode

The agent calls tools when it should ask a clarification question.

3. RAG Agent

Problem

The answer depends on knowledge outside the model's context.

Structure

Question -> Retrieve Documents -> Rank Context -> Generate Answer -> Cite Sources -> Evaluate Faithfulness

Use For

• internal knowledge assistants
• policy lookup
• product documentation
• research support

Failure Mode

The model answers confidently even when retrieval found the wrong documents.

4. Memory Agent

Problem

The agent needs stable context across tasks or sessions.

Structure

User Event -> Memory Write Policy -> Memory Store -> Retrieval Policy -> Current Context

Use For

• user preferences
• project facts
• long-running workflows
• shared team context

Failure Mode

The memory store becomes a pile of stale, sensitive, or irrelevant text.

5. Planner-Executor

Problem

The task has multiple steps and needs explicit control.

Structure

Task -> Plan -> Execute Step -> Observe -> Continue Or Stop -> Final Output

Use For

• research workflows
• report generation
• data analysis
• multi-step tool use

Failure Mode

The plan is generated but never checked against reality.

6. Reviewer-Evaluator

Problem

You need to catch errors before users see the final output.

Structure

Draft -> Rubric -> Review -> Pass Or Revise -> Final

Use For

• content generation
• code review support
• high-stakes summaries
• structured reports

Failure Mode

The evaluator rewards style instead of correctness.

7. Router-Specialist

Problem

Different task types require different prompts, tools, or policies.

Structure

Input -> Router -> Specialist Agent -> Specialist Output -> Final Formatter

Use For

• customer support categories
• domain workflows
• mixed internal tools
• enterprise automation

Failure Mode

The router is not evaluated, so requests go to the wrong specialist.

8. Supervisor Multi-Agent

Problem

Several specialist agents must collaborate under one owner.

Structure

User Task -> Supervisor -> Specialists -> Shared State -> Evaluator -> Final Answer

Use For

• agent colonies
• domain teams
• research teams
• operations workflows

Failure Mode

Agents debate forever or produce conflicting answers without resolution.

9. Human Approval Gate

Problem

Some actions are too risky to automate fully.

Structure

Proposed Action -> Risk Classifier -> Approval Request -> Execute Or Refuse -> Audit Log

Use For

• sending emails
• deleting data
• financial actions
• medical workflows
• production system changes

Failure Mode

The approval prompt is vague, so the human cannot make an informed decision.

10. Evaluation Harness

Problem

You need to know whether a change made the agent better or worse.

Structure

Eval Cases -> Agent Run -> Scoring -> Report -> Release Decision

Use For

• prompt changes
• tool changes
• memory changes
• model upgrades
• production regression testing

Failure Mode

The eval set only contains happy paths and misses the cases users actually care about.