Workflow
Design an Agent Architecture Before Writing Code
A one-page design process that prevents the most expensive agent rebuilds.
2 min readupdated 2026-08-01
/ quick answer
Specify the five architectural layers first — model, loop, tools, memory, policy — and only then choose a framework. A one-page design process that prevents the most expensive agent rebuilds.
A one-page design process that prevents the most expensive agent rebuilds. The problem it solves: Teams start with a framework and a prompt, then discover three weeks in that the agent has no memory strategy, no failure path and no cost ceiling. Specify the five architectural layers first — model, loop, tools, memory, policy — and only then choose a framework. It runs in 7 steps, starting with write the job in one sentence: input, output, success criteria, hard limits. This workflow node is part of the Onexial knowledge graph and links to related concepts, workflows and tools below.
Problem
Teams start with a framework and a prompt, then discover three weeks in that the agent has no memory strategy, no failure path and no cost ceiling.
Solution
Specify the five architectural layers first — model, loop, tools, memory, policy — and only then choose a framework.
Steps
- 01Write the job in one sentence: input, output, success criteria, hard limits.
- 02Choose the loop: ReAct for open-ended tasks, plan-and-execute for repeatable ones.
- 03List tools with read/write classification. Every write tool needs an approval rule.
- 04Pick a memory strategy: none, per-session, or persistent vector/SQL store.
- 05Set the policy boundary: max steps, max spend, forbidden actions, escalation path.
- 06Define 10 eval cases before building. These become your regression suite.
- 07Only now pick the framework (LangGraph, CrewAI, AI SDK, custom loop).
Tools Used
Prompts Used
Related Dictionary
/ frequently asked
What does the Design an Agent Architecture Before Writing Code workflow do?
Specify the five architectural layers first — model, loop, tools, memory, policy — and only then choose a framework.
What problem does Design an Agent Architecture Before Writing Code solve?
Teams start with a framework and a prompt, then discover three weeks in that the agent has no memory strategy, no failure path and no cost ceiling.
How many steps does Design an Agent Architecture Before Writing Code take?
7 steps. It starts with write the job in one sentence: input, output, success criteria, hard limits. and ends with only now pick the framework (langgraph, crewai, ai sdk, custom loop)..
Which tools does Design an Agent Architecture Before Writing Code need?
It uses agent-architecture-stack — each linked below with its own node.
↳ connected nodes
Dictionary↳ linked
Agent Architecture
Agent architecture is the structural blueprint of an AI agent: the model, the planning loop, the tools, the memory layer and the guardrails that decide how it acts.
Dictionary↳ linked
Agent Planning (ReAct, Plan-and-Execute)
Agent planning is how an AI agent decides its next step — reactively step-by-step (ReAct) or by drafting a full plan up front (plan-and-execute).
Dictionary↳ linked
Agent Memory
Persistent context that lets agents retain preferences, decisions, and prior work.
Tool Stack↳ linked
Agent Architecture Stack
The minimum tooling to design, run and observe a production agent.
Prompt↳ linked
Agent Architecture Spec Prompt
Turns a fuzzy agent idea into a reviewable five-layer architecture spec.
Dictionary↳ linked
Multi-Agent System
A multi-agent system splits a job across several specialised AI agents that coordinate through a shared plan, message bus or orchestrator.
Comparison↳ linked
Single Agent vs Multi-Agent System
One well-equipped agent beats a crowd for most jobs; multi-agent wins on genuinely separable, parallel work.
Comparison↳ linked
LangGraph vs CrewAI vs AI SDK for Agents
LangGraph for stateful control, CrewAI for fast role-based teams, AI SDK for shipping agents inside a product.
Use Case↳ linked
SaaS Cuts First-Response Time from 6h to 4min with an Agent
A 12-person SaaS shipped a tier-1 support agent and kept humans on exceptions only.