Most enterprise AI work should be a workflow, not an agent. That is also Anthropic's guidance. Their definition is short: workflows are systems where models and tools are orchestrated through predefined code paths, and agents are systems where the model directs its own process and tool use. Their advice follows from it: start with the simplest solution, and add complexity only when it demonstrably improves outcomes.
For many tasks the simplest solution is a single model call with retrieval and a few examples. The rest of this article is about what to do when that is not enough.
The decision
Ask two questions. Can you write down the steps in advance? Can you tolerate the model deciding what happens next?
| Situation | Choose | Why |
|---|---|---|
| Steps are known and repeat | Workflow | Predictable cost and latency. Each step can be tested alone. |
| Input types differ, but each type has a known handling | Workflow with routing | Specialized prompts beat one prompt trying to handle everything. |
| The number of steps depends on what the model finds | Agent, bounded | A fixed path cannot be written. Add a step limit and checkpoints. |
| The action has financial, legal, or customer consequence | Either, with a human gate | The architecture does not remove the need for a decision-maker. |
Anthropic is direct about the cost of agents: higher cost, and the risk of compounding errors. They recommend testing in sandboxes, getting ground truth from the environment at each step, and including stopping conditions such as a maximum number of iterations.
Five workflow patterns, with enterprise examples
Prompt chaining. Split a task into sequential steps and check the output between them. An invoice is extracted, validated against business rules in code, then drafted into a posting. Each step is small enough to test.
Routing. Classify the input, then send it to a specialized handler. Claims, tickets, and documents arrive in categories that need different treatment. Anthropic also notes the cost use: easy cases go to a smaller model, hard cases to a more capable one.
Parallelization. Two forms. Sectioning runs independent subtasks at once, such as checking separate clauses of one contract. Voting runs the same check several times where confidence matters, such as flagging high-risk content.
Orchestrator-workers. A central model breaks the task down, delegates to workers, and combines the results. Use it when the subtasks cannot be predicted from the task type, such as investigating an alert across several systems.
Evaluator-optimizer. One call drafts, another critiques, and the loop repeats. It works when criteria are clear and critique measurably improves the output, such as checking a regulatory narrative against a checklist.
- Inputdocument, alert, request
- Routesmall model classifies
- Workersrun in parallel
- Combineone model assembles
- Decideperson approves
- Recordinputs, outputs, edits
Multi-agent systems have a price
Anthropic's research team reports that agents typically use about four times the tokens of a chat interaction, and multi-agent systems about fifteen times. They found multi-agent designs suit breadth-first work that parallelizes, and suit tasks that need shared context less. That rules out many workflows people reach for them to solve. Use several agents when subtasks are independent and the value of the task pays for the tokens.
Context is a budget
Anthropic describes context engineering as curating the smallest set of high-signal tokens that gets the outcome you want. Recall degrades as context grows, so more context is not free. Four practices follow.
- Load just in time. Keep lightweight identifiers such as file paths, record IDs, and stored queries, and let the system fetch the data when it needs it. Folder names, timestamps, and naming give the model signals about relevance.
- Compact long sessions. Summarize a conversation nearing its limit and continue in a fresh window. Clearing old tool results is a lighter form of the same idea.
- Write notes outside the window. Persistent notes let a long task keep its progress across many steps.
- Give subtasks clean windows. Sub-agents do detailed work in their own context and return a condensed summary to the lead.
Agent Skills apply the same principle to instructions. Each skill costs roughly 100 tokens of metadata until it is triggered, and its full instructions and files load only when needed. That lets you install many specialist procedures without paying for all of them on every request.
Tools are interface design
Anthropic's team reports spending more effort on tools than on the overall prompt. Their guidance reads like interface design for a junior colleague.
- Build a few tools for high-impact workflows and consolidate multi-step operations. More tools do not always lead to better outcomes.
- Return meaningful context, such as names instead of opaque IDs, and offer concise and detailed response modes.
- Keep outputs token-efficient with pagination, filtering, and truncation. Write error messages that say what to do next.
- Make mistakes harder to commit. In one example, requiring absolute file paths removed a class of errors that relative paths caused.
What we do by default
We start with a workflow and a human decision at the end. We add an agent step only where the path genuinely cannot be known, and we bound it with a step limit and a log. We write the evaluation before the system, which is the subject of the next article. Anthropic Academy offers a course, Building with the Claude API, covering prompting, tool use, retrieval, agents, and MCP, plus shorter courses on subagents and agent skills. They are a reasonable shared vocabulary for a team new to this.
This is a reference design. It describes how we would build the system and cites Anthropic's documentation for platform facts. It is not a client case study, and the figures it uses are list prices or worked examples, not results.
Have a process like this?
Bring it to a 45-minute working session.