When a Workflow Beats a Loop
This content was developed with AI assistance and is regularly reviewed for accuracy.
Snapshot: October 2026. Model names, prices, and product features change monthly; What's New in AI is the maintained chronology.
Every other recipe in this section builds a loop. This one helps you decide not to. An agent loop chooses its own steps toward a goal, which is exactly what you want when the work changes shape every run and exactly what you do not want when the work must be approved, audited, or repeated identically. Most tasks people want to loop are better served by something simpler: a fixed workflow, a workflow with one AI step, or a single scheduled prompt. Thirty minutes with this recipe can save you weeks of babysitting a loop that never needed to exist.
Last tested with Claude Code Routines, Claude scheduled tasks, and a no-code workflow tool in October 2026.
What you'll accomplish
- Place a task on the automation ladder and pick the lowest rung that works
- Run five tests that separate workflow work from loop work
- Compare the cost of each option, including the human review time
- Set a revisit trigger so you promote the task to a loop only when it earns it
Prerequisites
- One real task you are considering automating, with a week of examples
- What Is Loop Engineering? and Workflows vs Agents read first
- 30 minutes
Why the negative case matters
The evidence for caution is consistent. Gartner's standing forecast, published in June 2025, is that over 40% of agentic AI projects will be canceled by the end of 2027 because of escalating costs, unclear business value, or inadequate risk controls; in the same release it estimated that only about 130 of the thousands of vendors marketing "agents" were selling the real thing. Menlo Ventures' December 2025 enterprise survey found only 16% of enterprise deployments qualified as true agents that plan, act, and adapt; the rest were copilots and workflows. Reviewers of 2026's consumer agents agree on where they break: website logins, long unattended runs, and anything that must be repeatable and auditable (reported).
Governing agents also costs real effort. Anthropic reported running about 30,000 agents concurrently on its most-used internal platform with 100% of actions monitored before execution, which is excellent practice and an investment few teams can match. And the year's signature failure, the PocketOS database deletion in April 2026, came from an agent on a routine task improvising with a key it found, the kind of improvisation a fixed workflow cannot do.
None of this says loops are bad. It says the loop has to beat the simpler option on the task in front of you.
The automation ladder
Four rungs cover almost everything. Climb only as high as the task requires.
| Rung | What it is | Example |
|---|---|---|
| 1. A brief you run yourself | A saved, reusable brief or playbook you trigger by hand | Drafting the monthly board summary |
| 2. A one-turn scheduled prompt | One prompt on a schedule, one result delivered, no iteration | A Monday digest of last week's support themes |
| 3. A workflow, optionally with an AI step | Fixed steps and branches; a model makes one bounded judgment such as classify or extract | Routing inbound inquiries to the right queue |
| 4. An agent loop | The agent finds work, acts, checks, and decides what's next until a stop condition | Fixing flaky tests from a queue overnight |
Rung 2 is the most underused. If the task is "every Monday, read X and tell me Y," that is a scheduled prompt, not a loop: there is nothing to check, retry, or decide between runs. Claude's scheduled tasks, a Routine, or the weekly report playbook covers it.
Step 1: Write the task down as steps
Goal: Find out whether the path is actually fixed.
Action: Take five real examples from last week and write the steps you followed for each, including every branch ("if the amount is over $5,000, send to the director"). Then compare the five lists.
If the lists are the same apart from one judgment call, you have a workflow with an AI step. If they are identical, you have a plain workflow. If each one took a different route that you could not have written down in advance, you may have loop work.
Checkpoint: You can say, from evidence, whether the path changes from run to run.
Step 2: Run the five tests
Goal: Score the task against the questions that decide the rung.
Action: Answer each about the task as it is today, not as you hope it will be.
| Test | Points to a workflow or scheduled prompt | Points to a loop |
|---|---|---|
| Path | Every step and branch can be written down | The path changes each run |
| Approval and audit | Someone must approve or audit every run | Review by exception is enough |
| Repetition | Results must be identical for identical input | Variation is acceptable, even useful |
| Reversibility | A wrong action is costly or irreversible | A wrong action is cheap to undo, such as a pull request nobody merged |
| Checkability | Success is obvious from the output format | Success needs an objective check you can automate, and you can write the stop condition |
Count the answers. Three or more in the left column means a workflow or scheduled prompt. A loop needs the right column on path, reversibility, and checkability at minimum; if you cannot write the stop condition and the success check, you are not ready to write the loop.
Checkpoint: You have a scored table and a provisional rung.
Step 3: Compare the cost
Goal: Price both options honestly, including people's time.
Action: Estimate a month of each. For the loop, use tokens per iteration, iterations per run, and runs per month, as in Loop Guardrails and Safety, and remember that agent teams can use about seven times the tokens of a single session. For the workflow, use the tool's per-task or per-step pricing plus any AI-step usage. Then add the line most estimates leave out: the human time to review output, investigate surprises, and maintain the setup.
A loop that saves four hours of work but needs three hours of review and debugging is not a win. A workflow that handles 80% of cases and sends the rest to a person often is. Evaluate and Measure has the full business-case method.
Checkpoint: You have a monthly cost for each option, with review time included.
Step 4: Check approval and audit requirements
Goal: Make sure the option you pick can satisfy whoever signs off.
Action: Ask who needs to approve the work and who may audit it later. Regulated decisions, money movement, customer communications, and anything affecting a person's health, legal rights, finances, or livelihood usually need a named human approval and a record. Some vendor terms now require this: Anthropic's usage policy effective November 12, 2026 calls for a qualified human reviewer and notice to the affected person in those high-stakes uses.
A workflow is auditable by construction: every branch was written down before it ran, and tools such as Relay.app are built around a step that pauses until a person approves. A loop can include approvals too, but each one interrupts the autonomy you built it for.
Checkpoint: You know which approvals the task needs and whether your chosen rung provides them cleanly.
Step 5: Build the smallest version that passes
Goal: Ship the lowest rung that meets the tests, then measure it.
Action: If the answer is a workflow with an AI step, the inquiry router tutorial is the template: a fixed flow, one classification step, and a human queue for low-confidence cases. If the answer is a scheduled prompt, write the brief and schedule it. If the answer really is a loop, go to The Overnight Coding Loop or The Research and Monitoring Loop.
Hybrids are common and healthy. A workflow can call an agent for one hard step and keep the approval gates around it. A loop can sit inside deterministic checks, such as CI that must pass before anything counts as done.
Checkpoint: The simplest working version is live and logging results.
Step 6: Set a revisit trigger
Goal: Promote the task to a loop only when the evidence says so.
Action: Write down the condition that would justify climbing a rung, and check it quarterly. Good triggers are measurable: the share of cases going to the human queue stays above a threshold you set, the steps keep changing so the workflow needs weekly edits, or a new model makes the hard judgment reliable enough to automate. Re-examine the decision with each new model, as Anthropic recommends for harnesses; the reverse is also true, and a model upgrade sometimes makes a workflow's AI step good enough that the loop you planned is unnecessary.
Checkpoint: The trigger and the review date are on a calendar.
A worked example
A finance operations lead had two candidate loops.
Expense-report routing. Every report follows the same policy: check receipts, apply limits, send anything over threshold to a director. Five examples produced identical step lists apart from reading messy receipts. Every run needs an auditable record and approvals above the limit. Score: four of five tests on the workflow side. Result: a workflow with one AI step that extracts receipt fields, and a human approval for exceptions. Cost: a fraction of the loop estimate, and the auditors were satisfied on day one.
Flaky-test cleanup. Each failure has a different cause, a wrong fix is a pull request nobody merges, and CI provides an objective check. Score: loop side on path, reversibility, and checkability. Result: a loop on the overnight-coding pattern, with a runtime budget and human merges.
Same team, same week, two different rungs. That is the point of the recipe.
Troubleshooting
- A vendor calls its workflow an agent. Ask whether it chooses its own steps or follows a configured path. Either can be fine; just do not pay agent prices or accept agent risk for a workflow.
- The team wants a loop because loops are interesting. Run the five tests together and price both options; the numbers usually settle it.
- The workflow keeps breaking on new cases. That is the revisit trigger firing. Add an AI step for the messy input before you jump to a loop.
Key takeaways
- Climb the ladder only as high as the task needs: brief, scheduled prompt, workflow with an AI step, then loop
- Fixed paths, per-run approvals, identical repetition, and costly mistakes all point to workflows
- A loop needs a changing path, cheap reversibility, and a success check you can automate and a stop condition you can write
- Price review and maintenance time along with tokens and tool fees
- Set a measurable trigger for promoting a task, and revisit it when models change
Next steps
Continue to: Evals for Agents to build the checks any loop you do build will need.
- Want the full workflow-versus-agent decision rule? Read Workflows vs Agents
- Building the workflow instead? Follow Tutorial: Build an Inquiry Router With an AI Classifier
- Making the business case to leadership? Use Evaluate and Measure: Tools, ROI, Impact