Agent Developer
A reference skill for designing and building AI agents. It carries no deep knowledge itself — it routes you to focused reference files, each loaded only when you need it.
Based on "Patterns for Building AI Agents" and "Principles of Building AI Agents" (Bhagwat & Gienow, 2025), plus the twelve-factor-agents methodology.
How to use this skill
Starting a new agent? Read references/architecture.md first. It walks you through capability mapping and choosing an architecture (Single Agent, Router + Specialists, Coordinator + Workers, or Pipeline). The architecture decision determines which other references matter.
Working on an existing agent? Jump straight to the topic section below that matches the subsystem you are touching, and read its reference file.
Each reference file is dense, agent-readable knowledge meant to be consulted while building — frameworks, decision trees, matrices, and checklists.
Architecture
Choosing the overall shape of the agent system: how many agents, how they relate, and how the design evolves.
- Architecture & evolution: Capability mapping, the architecture selection matrix, the human-in-the-loop framework, and dynamic-configuration signals. Read references/architecture.md
Prompting & model selection
The model you pick and the prompt you write are the foundation of every agent.
- Prompts & models: Model-selection matrix, the five prompt-architecture layers, seed-crystal bootstrapping, few-shot guidelines, and the production checklist. Read references/prompting.md
Tools
How the agent acts on the world. Tool design is the single most important step in building an agent.
- Tool design: Operation decomposition, the tool-schema template, the third-party integration map, the MCP decision tree, performance targets, and the structured error shape. Read references/tools.md
Memory
How the agent remembers across turns and sessions.
- Memory architecture: The three-layer memory model, the agent-type-to-layer matrix, the working-memory schema, semantic-recall configuration, and memory processors. Read references/memory.md
Context engineering
How context flows through the agent and how to keep the context window healthy.
- Context strategy: Parallelization rules, context-sharing strategies, the five context failure modes, compression strategy, and the error-feedback loop. Read references/context.md
Workflows
When pure agentic looping is too unpredictable, give the agent explicit structure.
- Graph workflows: The four workflow primitives, suspend/resume, streaming, OpenTelemetry tracing, and workflow composition. Read references/workflows.md
RAG
When the agent needs to retrieve knowledge from a corpus — and how to decide whether RAG is even the right tool.
- RAG pipeline: The RAG decision tree, chunking / embedding / vector-DB matrices, retrieval tuning, the pipeline architecture, and the quality checklist. Read references/rag.md
Multi-agent systems
When one agent is not enough and the work needs a team of collaborating agents.
- Multi-agent design: Single-vs-multi criteria, supervision patterns, organizational design, control flow, the A2A protocol, and failure handling. Read references/multi-agent.md
Evaluation
Define what "good" looks like and measure it continuously.
- Eval system: The failure-mode taxonomy, business metrics, eval-suite structure, SME labeling, and the production-data pipeline. Read references/evaluation.md
Production reliability
Cross-cutting principles for making an agent reliable, debuggable, and operable in production.
- Twelve-factor agents: The twelve factors — natural-language-to-tool-calls, owning your prompts and context window, treating tools as structured outputs, unified state, launch/pause/resume, human contact as tool calls, owning control flow, compacting errors, small focused agents, triggering from anywhere, and the stateless-reducer model. Read references/twelve-factor.md
1---2name: agent-developer3description: Designs and builds AI agents. Use when working on agent architecture, prompts, model selection, tool design, MCP, memory, context engineering, RAG, graph workflows, multi-agent systems, evaluation, or production reliability.4---56# Agent Developer78A reference skill for designing and building AI agents. It carries no deep knowledge itself — it routes you to focused reference files, each loaded only when you need it.910Based on "Patterns for Building AI Agents" and "Principles of Building AI Agents" (Bhagwat & Gienow, 2025), plus the twelve-factor-agents methodology.1112## How to use this skill1314**Starting a new agent?** Read [references/architecture.md](references/architecture.md) first. It walks you through capability mapping and choosing an architecture (Single Agent, Router + Specialists, Coordinator + Workers, or Pipeline). The architecture decision determines which other references matter.1516**Working on an existing agent?** Jump straight to the topic section below that matches the subsystem you are touching, and read its reference file.1718Each reference file is dense, agent-readable knowledge meant to be consulted *while* building — frameworks, decision trees, matrices, and checklists.1920## Architecture2122Choosing the overall shape of the agent system: how many agents, how they relate, and how the design evolves.2324- **Architecture & evolution**: Capability mapping, the architecture selection matrix, the human-in-the-loop framework, and dynamic-configuration signals. Read [references/architecture.md](references/architecture.md)2526## Prompting & model selection2728The model you pick and the prompt you write are the foundation of every agent.2930- **Prompts & models**: Model-selection matrix, the five prompt-architecture layers, seed-crystal bootstrapping, few-shot guidelines, and the production checklist. Read [references/prompting.md](references/prompting.md)3132## Tools3334How the agent acts on the world. Tool design is the single most important step in building an agent.3536- **Tool design**: Operation decomposition, the tool-schema template, the third-party integration map, the MCP decision tree, performance targets, and the structured error shape. Read [references/tools.md](references/tools.md)3738## Memory3940How the agent remembers across turns and sessions.4142- **Memory architecture**: The three-layer memory model, the agent-type-to-layer matrix, the working-memory schema, semantic-recall configuration, and memory processors. Read [references/memory.md](references/memory.md)4344## Context engineering4546How context flows through the agent and how to keep the context window healthy.4748- **Context strategy**: Parallelization rules, context-sharing strategies, the five context failure modes, compression strategy, and the error-feedback loop. Read [references/context.md](references/context.md)4950## Workflows5152When pure agentic looping is too unpredictable, give the agent explicit structure.5354- **Graph workflows**: The four workflow primitives, suspend/resume, streaming, OpenTelemetry tracing, and workflow composition. Read [references/workflows.md](references/workflows.md)5556## RAG5758When the agent needs to retrieve knowledge from a corpus — and how to decide whether RAG is even the right tool.5960- **RAG pipeline**: The RAG decision tree, chunking / embedding / vector-DB matrices, retrieval tuning, the pipeline architecture, and the quality checklist. Read [references/rag.md](references/rag.md)6162## Multi-agent systems6364When one agent is not enough and the work needs a team of collaborating agents.6566- **Multi-agent design**: Single-vs-multi criteria, supervision patterns, organizational design, control flow, the A2A protocol, and failure handling. Read [references/multi-agent.md](references/multi-agent.md)6768## Evaluation6970Define what "good" looks like and measure it continuously.7172- **Eval system**: The failure-mode taxonomy, business metrics, eval-suite structure, SME labeling, and the production-data pipeline. Read [references/evaluation.md](references/evaluation.md)7374## Production reliability7576Cross-cutting principles for making an agent reliable, debuggable, and operable in production.7778- **Twelve-factor agents**: The twelve factors — natural-language-to-tool-calls, owning your prompts and context window, treating tools as structured outputs, unified state, launch/pause/resume, human contact as tool calls, owning control flow, compacting errors, small focused agents, triggering from anywhere, and the stateless-reducer model. Read [references/twelve-factor.md](references/twelve-factor.md)
IKatsuba