The Three Architectural Patterns: Augmented LLM, Workflow, Agent
CoreSelect among the augmented LLM, workflow, and agentic architectural patterns · Difficulty 3/5
Explanation
Anthropic frames Claude systems along a complexity gradient. Knowing when each pattern fits -- and picking the simplest one that satisfies the constraints -- is the architect's core judgment call.
The Complexity Gradient
| Pattern | What it is | Fits when |
|---|---|---|
| Augmented LLM | A single model call enhanced with retrieval, tools, and memory | The base building block; one well-scoped step |
| Workflow | LLM calls and tools orchestrated through predefined code paths | Task decomposes into known, fixed steps; predictability and testability matter |
| Agentic | The LLM dynamically directs its own steps and tool use | Open-ended tasks where the steps can't be enumerated in advance |
The Augmented LLM Is the Atomic Unit
The augmented LLM -- model + retrieval + tools + memory -- is the atomic unit that most workflows and agents are compositions of. It is a single, well-scoped model call that has been given the context (via retrieval), the capabilities (via tools), and the continuity (via memory) it needs to complete one step reliably.
What Makes a System an Agent
An agent proper runs an open-ended loop: plan -> call tool -> observe real environment feedback -> repeat until done or a stop condition trips. Agents trade latency, cost, and predictability for the ability to handle tasks whose path can't be pre-scripted. The defining feature is that the *LLM itself* -- not the surrounding code -- decides what happens next at each step.
Choosing the Simplest Fitting Pattern
Anthropic's recurring guidance is disciplined restraint: find the simplest solution that works, and add complexity (autonomy, multi-agent orchestration) only when the task genuinely benefits. A single augmented LLM call should be the default hypothesis; escalate to a workflow only when the task has multiple fixed steps, and to an agent only when the steps genuinely cannot be enumerated ahead of time.
Key Takeaways
- Three patterns on a complexity gradient: augmented LLM (base unit), workflow (predefined code paths), agent (LLM-directed loop)
- The augmented LLM -- model + retrieval + tools + memory -- is the atomic unit that workflows and agents compose
- Workflows fit known, fixed steps where predictability and testability matter
- Agents run plan -> call tool -> observe -> repeat, trading latency/cost/predictability for handling open-ended tasks
- Always start from the simplest pattern that satisfies the requirement; add complexity only when the task genuinely benefits
Glossary Terms
Related Concepts
Workflow Composition Patterns & the Orchestrator-Workers Trap
Five workflow composition patterns: prompt chaining, routing, parallelization, orchestrator-workers, evaluator-optimizer
Manager/Subagent Orchestration & Context Isolation
Manager/orchestrator coordinates specialized subagents, each with its own context window