git:20260424.1fce5c1 to git:20260430.c92465c
11 added, 70 removed. Audit A to A.
---
name: agentic-patterns
description: |
Use when designing or analyzing multi-agent architectures — covers core patterns (sequential, hierarchical,
parallel, event-driven), anti-patterns, and the five-layer model (Cognition → Context → Interaction →
Runtime → Governance) for separating concerns in production-grade agentic systems.
NOT for reviewing existing agentic code (use review-methodology) or framework selection (use the frameworks skill).
portability: portable
---
# Agentic Patterns
Reference guide for proven agentic architecture patterns and common anti-patterns to avoid.
## Core Patterns
- ### Sequential Pattern
- Agents execute in fixed order. Each agent's output feeds the next.
-
- **Use when:** Linear workflow, clear dependencies, predictable stages.
-
- **Example:** Research → Analyze → Draft → Review
-
- **Tradeoffs:** Simple coordination, but bottlenecks if any agent is slow.
-
- ### Hierarchical Pattern
- Parent agent delegates to specialized child agents. Parent aggregates results.
-
- **Use when:** Complex tasks decomposable into subtasks, need centralized coordination.
-
- **Example:** Project manager agent delegates to designer, developer, QA agents.
-
- **Tradeoffs:** Clear accountability, scales well, but single point of failure at parent.
-
- ### Collaborative Pattern
- Peer agents work together on shared problem. No fixed hierarchy.
-
- **Use when:** Problem benefits from diverse perspectives, no clear lead agent.
-
- **Example:** Multiple specialized reviewers collaborating on code review.
-
- **Tradeoffs:** Rich outputs, but requires conflict resolution and consensus mechanisms.
-
- ### Autonomous Pattern
- Agents operate independently with minimal coordination.
-
- **Use when:** Tasks are fully independent, parallel execution needed.
-
- **Example:** Multiple monitoring agents checking different services.
-
- **Tradeoffs:** Maximum parallelism, but no shared learning or coordination.
-
- ### Human-in-the-Loop Pattern
- Human approval gates at critical decision points.
-
- **Use when:** High-stakes decisions, regulatory requirements, learning from human feedback.
-
- **Example:** Agent drafts report, human approves before sending.
-
- **Tradeoffs:** Adds latency, but ensures safety and quality.
-
- ### ReAct Pattern
- Reason → Act → Observe cycle. Agent reasons about next action, executes it, observes result, repeats.
-
- **Use when:** Dynamic environments, need adaptive behavior based on feedback.
-
- **Example:** Debugging agent tries fixes, observes test results, adapts approach.
-
- **Tradeoffs:** Flexible and adaptive, but can be inefficient if reasoning is expensive.
-
- ### Plan-Execute Pattern
- Agent creates full plan upfront, then executes all steps.
-
- **Use when:** Environment is predictable, planning overhead is justified.
-
- **Example:** Multi-step data pipeline with known transformations.
-
- **Tradeoffs:** Efficient execution, but brittle if environment changes.
-
- ### Reflection Pattern
- Agent reviews its own outputs and iteratively improves them.
-
- **Use when:** Quality matters more than speed, self-improvement is valuable.
-
- **Example:** Writer agent drafts, critiques, and refines output.
+ | Pattern | Use when | Example | Tradeoffs |
+ |---------|----------|---------|-----------|
+ | **Sequential** | Linear workflow, clear dependencies, predictable stages | Research → Analyze → Draft → Review | Simple coordination; bottlenecks if any agent is slow |
+ | **Hierarchical** | Complex task decomposable into subtasks; central coordination | PM agent delegates to designer / dev / QA | Clear accountability, scales; single point of failure at parent |
+ | **Collaborative** | Diverse perspectives matter; no clear lead | Multiple reviewers collaborating on code review | Rich outputs; needs conflict resolution + consensus |
+ | **Autonomous** | Tasks are fully independent; parallel execution | Multiple monitoring agents checking different services | Max parallelism; no shared learning or coordination |
+ | **Human-in-the-Loop** | High-stakes decisions, regulatory gates, human feedback | Agent drafts report; human approves before sending | Adds latency; ensures safety and quality |
+ | **ReAct** | Dynamic environments; need adaptive behavior on feedback | Debug agent: try fix → observe tests → adapt | Flexible; high token/latency cost for reasoning |
+ | **Plan-Execute** | Predictable environment; planning overhead justified | Multi-step pipeline with known transformations | Efficient; brittle if environment changes |
+ | **Reflection** | Quality > speed; self-improvement valuable | Writer drafts → critiques → refines | Higher quality; more tokens and latency |
- **Tradeoffs:** Higher quality outputs, but increased token usage and latency.
+ See `refs/pattern-catalog.md` for full descriptions and implementation notes per pattern.
## Anti-Patterns Quick Reference
1. **God Agent:** Single agent doing everything. Split responsibilities.
2. **Tight Coupling:** Agents depend on internal implementation details. Use interfaces.
3. **Missing Guardrails:** No constraints on agent actions. Add validation and limits.
4. **Deep Nesting:** Agents calling agents calling agents. Flatten hierarchy.
5. **No Observability:** Can't see what agents are doing. Add logging and tracing.
6. **Sequential Bottleneck:** Everything waits for slowest agent. Add parallelism.
7. **Context Bloat:** Passing entire history to every agent. Compress and filter.
8. **Redundant Agents:** Multiple agents doing same thing. Consolidate.
9. **Hardcoded Prompts:** Prompts baked into code. Externalize and version.
10. **Missing Timeouts:** Agents can run forever. Add time and resource limits.
## Pattern Selection Decision Tree
```
Is task decomposable into subtasks?
├─ YES: Is there a clear lead/coordinator role?
│ ├─ YES: Use Hierarchical
│ └─ NO: Use Collaborative
└─ NO: Does task require adaptation based on results?
├─ YES: Can you plan all steps upfront?
│ ├─ YES: Use Plan-Execute
│ └─ NO: Use ReAct
└─ NO: Do you have multiple independent tasks?
├─ YES: Use Autonomous
└─ NO: Use Sequential
```
Add Human-in-the-Loop gates for any high-stakes decisions.
Add Reflection layer when quality is critical.
## Combining Patterns
Patterns are composable:
- **Hierarchical + ReAct:** Parent plans, children adapt execution
- **Sequential + Reflection:** Each stage includes self-review
- **Collaborative + Human-in-Loop:** Peer review with human arbitration
- **Autonomous + Hierarchical:** Independent teams with team leads
## Pattern Anti-Affinities
Avoid these combinations:
- **Plan-Execute + ReAct:** Contradictory - either plan upfront or adapt dynamically
- **Autonomous + Tight Coordination:** Defeats the purpose of autonomy
- **Deep Hierarchies + Reflection:** Exponential token cost
## Quick Start
1. Identify if task is decomposable (subtasks?) or atomic
2. Identify if execution is predictable (plan?) or dynamic (adapt?)
3. Identify if agents should coordinate (hierarchy/collaboration) or work independently
4. Check anti-patterns list against current design
5. See refs/pattern-catalog.md for detailed examples
## References
- `refs/pattern-catalog.md` - Detailed pattern descriptions with implementation examples
- `refs/anti-patterns-design.md` - Design anti-patterns (God Agent, Tight Coupling, Missing Guardrails, Deep Nesting, No Observability)
- `refs/anti-patterns-operational.md` - Operational anti-patterns (Sequential Bottleneck, Context Bloat, Redundant Agents, Hardcoded Prompts, Missing Timeouts)