AI Application Spectrum
Summary: LLM applications sit on a spectrum from a single augmented call to fully autonomous agents, and the tier determines how much control the developer keeps over what happens.
At a glance​
| Question | Answer |
|---|---|
| What is it? | A way to classify LLM applications by how much autonomy the model has |
| Why does it matter? | The right tier depends on the task; more autonomy isn't always better |
| When do I use it? | When deciding how much structure to build around an LLM call |
How it works​
- Augmented LLM call: a single call plus tools, retrieval, or memory. No autonomy — the developer controls the flow entirely.
- Workflow: the developer defines fixed steps, orchestrated through predefined code paths; the LLM fills in each step. See LLM Workflow Patterns.
- Agent: the LLM dynamically directs its own process and tool usage, maintaining control over how it accomplishes the task. See Agents.
- Multi-agent system: several agents with different roles coordinate, sometimes with one orchestrating the others. Anthropic's own taxonomy formally recognizes only two categories — workflows and agents — and treats multi-agent setups as multiple agents coordinating, not a distinct third tier.
When to use each tier​
| Tier | Use when | Avoid when |
|---|---|---|
| Augmented LLM call | Task is a single, well-defined call — retrieval, one tool, or memory is enough | The task needs multiple dependent steps or a decision about what to do next |
| Workflow | Steps are known and fixed ahead of time, even if each step's content varies | Steps genuinely can't be predicted before the input arrives |
| Agent | Task is open-ended, the number of steps can't be hardcoded, and the environment is trusted enough for autonomy | The task is well-defined enough to fit a workflow — agents cost more and are harder to predict |
| Multi-agent | Distinct roles genuinely benefit from separation, and the coordination overhead is worth it | A single agent already handles the task — added coordination adds complexity without payoff |
Remember​
- More autonomy trades control for flexibility — pick the lowest tier that gets the job done reliably.
- A workflow is still deterministic in its steps; only the content of each step is generated.
- An agent's step count and tool calls aren't fixed in advance, which makes cost and behavior harder to predict.
- Multi-agent is agents at scale, not a separate architectural rung — this matches the industry-standard framing, not just a judgment call.
Related notes​
- LLM Workflow Patterns - the patterns used to implement the "workflow" tier
- LLM Workflow App Practices - how to build the "workflow" tier reliably
- Agents - the "agent" tier in depth
Sources​
- Building Effective Agents - Anthropic, Dec 2024