Memory Architecture for AI Agents: Sessions Knowledge and Durable Lessons
September 1, 2026
HP ZGX Nano 4 TB Review: The Enterprise-Minded GB10 Workstation
September 4, 2026
The most capable architecture is not automatically the best architecture. A fixed workflow, one tool-using agent and a multi-agent system trade predictability for flexibility at different rates. Start with the smallest design that can meet the job.
Deterministic workflow
A workflow routes inputs through known steps. Models may classify, extract or draft inside those steps, but application code determines the sequence.
Choose it when inputs and outputs are structured, the path is stable and failures must be easy to reproduce. Workflows are well suited to document pipelines, approval routing and transformations with objective validation.
Single agent
A single agent chooses which tool to use and how to adapt after each observation. It is useful when the path cannot be fully specified in advance: research, troubleshooting, code exploration or work across changing sources.
The trade-off is variability. The harness must limit tools, preserve state, enforce budgets and verify completion.
Multi-agent system
Multiple agents can divide work, contribute different tools or perform independent review. Anthropic’s multi-agent research architecture uses a lead agent and parallel specialists for breadth-first investigation. The same engineering report warns that the approach consumes substantially more tokens and is a poor fit when tasks have tight dependencies or require shared context.
Use multiple agents when parallel exploration, permission isolation or genuinely different expertise creates information a single agent would not obtain efficiently.
Decision matrix
| Condition | Workflow | Single agent | Multi-agent |
|---|---|---|---|
| Steps known in advance | Best fit | Usually unnecessary | Avoid |
| Dynamic tool choice | Limited | Strong fit | Possible |
| Independent parallel branches | Manual orchestration | Limited | Strong fit |
| Strict reproducibility | Strongest | Requires controls | Hardest |
| Cost and latency sensitivity | Lowest | Medium | Highest |
| Permission isolation | Code roles | Tool-level | Agent-level possible |
A staged path
Build a workflow first. Replace only the part that requires dynamic judgment with an agent. Add another agent only after evaluation shows a specific bottleneck that independent context or parallel work solves.
This progression produces a useful counterfactual: if the multi-agent version is better, you can compare it with the single-agent and workflow baselines. Without those baselines, architectural complexity becomes a belief.
Common design mistake
Do not assign personas and call the result multi-agent architecture. Roles need distinct responsibility, context, tools or verification value. Three agents reading the same prompt and voting may amplify the same blind spot.
Use What Is an AI Agent? for the base architecture, Multi-Agent Systems for collaboration patterns and How to Evaluate an AI Agent for evidence.
Primary sources
- Anthropic: Building effective agents
- Anthropic: Multi-agent research system
- OpenAI: A practical guide to building agents
Adaptation note: This article was informed by orchestration and collaboration concepts in AI Agents in Depth: Design Principles and Engineering Practice by Bojie Li and contributors, Apache License 2.0. It was independently structured and written for Stariy.com.
