Agent Workflows: Single-Step vs Multi-Step

Single-Step Workflows

A single-step workflow processes the task in one model call with no tool use. The agent receives input, generates output, and the interaction is complete. Classification, summarization, extraction, and formatting tasks all work well as single-step workflows. The agent does not need to search for information, compute anything, or modify external state, so the overhead of a multi-step reasoning loop provides no benefit.

Single-step workflows are the fastest and cheapest pattern because they involve exactly one model call. They are also the most predictable because the output is deterministic given the same input and model configuration. Use this pattern whenever the task can be completed with the information already present in the prompt. Adding unnecessary tool calls or reasoning steps to a task that does not need them wastes time, money, and increases the chance of errors.

Linear Multi-Step Workflows

Linear workflows execute a predetermined sequence of steps. The output of each step feeds into the next. A content creation workflow might follow this sequence: research the topic, outline the article, write each section, review for accuracy, format for publication. Each step completes before the next begins, and the workflow always follows the same path regardless of what it finds along the way.

The advantage of linear workflows is predictability. You know exactly how many steps the agent will take, what tools it will use, and what the intermediate outputs will look like. This makes testing, monitoring, and cost estimation straightforward. If step three always takes about 2,000 tokens and one tool call, you can predict the cost of the entire workflow with reasonable accuracy.

The disadvantage is rigidity. Linear workflows cannot adapt to unexpected results. If the research step finds that the topic is too broad to cover in a single article, the workflow continues with the same outline and writing steps regardless, producing a shallow article rather than suggesting a different approach. This rigidity is acceptable for well-understood tasks with predictable inputs but problematic for tasks with high variability.

Branching Workflows

Branching workflows include decision points where the agent evaluates intermediate results and chooses which path to follow. After classifying a customer support ticket, the agent might branch to different resolution workflows depending on the ticket category. A billing issue follows a different process than a technical support issue, which follows a different process than a feature request.

Branches can be conditional (based on a single criterion) or multi-way (based on classification into multiple categories). They can also be nested, with sub-branches within branches creating a decision tree that handles increasingly specific scenarios. The depth and breadth of branching should match the variability of the task. Over-branching creates unnecessary complexity. Under-branching forces diverse situations into generic workflows that do not handle any of them well.

Testing branching workflows requires testing every path, not just the most common one. A workflow with five binary decision points has 32 possible paths. If each branch is tested in isolation but the combinations are not tested together, edge cases at branch intersections can produce unexpected behavior. Comprehensive testing of branching workflows is significantly more expensive than testing linear workflows, which is a real cost that should factor into the design decision.

Autonomous Workflows

Autonomous workflows give the agent a goal without prescribing the steps. The agent decides what to do, in what order, using what tools, based on its own reasoning. This is the most powerful and most dangerous pattern. Powerful because the agent can handle novel situations, adapt to unexpected data, and find creative solutions that no prescribed workflow would have included. Dangerous because the agent might take inefficient paths, enter loops, make costly mistakes, or take actions with unintended consequences.

Autonomous workflows require strong guardrails to be production-safe. Turn limits prevent infinite loops by capping the maximum number of reasoning turns. Cost budgets cap the total spending on model calls and tool executions. Action approval gates require human approval before executing high-impact tools. Monitoring and alerting detect anomalous behavior (unusual tool call patterns, excessive retries, escalating costs) and can pause the agent for human review.

Most production agents use a hybrid approach: autonomous reasoning within a constrained set of tools and actions. The agent is free to decide how to accomplish the goal, but it can only use pre-approved tools, it cannot exceed defined budgets, and certain actions require human approval. This constrained autonomy provides most of the flexibility benefits with manageable risk.

Parallel Workflows

Parallel workflows split independent subtasks across multiple concurrent executions. If the agent needs to research five companies, it can research all five simultaneously rather than sequentially. Parallel execution reduces total completion time by a factor roughly equal to the number of parallel tasks, making it essential for time-sensitive operations that involve multiple independent data gathering or processing steps.

Implementing parallel workflows requires identifying which subtasks are truly independent. Two subtasks are independent if neither requires the output of the other. Researching company A and researching company B are independent. Researching a company and then analyzing the research results are dependent. Incorrectly parallelizing dependent subtasks produces errors because one subtask tries to use data that has not been generated yet.

Result aggregation is the final step of parallel workflows. After all parallel subtasks complete, their results must be combined into a coherent output. This aggregation step often requires a separate model call that synthesizes the individual results, resolves conflicts, and produces a unified response. The quality of aggregation depends on how well the individual results were structured and labeled, which is why parallel task design should include standardized output formats.

Choosing the Right Pattern

Start with the simplest pattern that could work. Single-step first, linear multi-step second, branching third, autonomous last. Each increase in complexity brings additional failure modes, testing requirements, and operational overhead. The right pattern is the simplest one that handles the required variability of the task without compromising quality.

Task variability is the primary driver of pattern selection. Tasks with low variability (same inputs, same process, same outputs) work well with linear workflows. Tasks with moderate variability (different input categories requiring different processes) need branching. Tasks with high variability (novel situations, unpredictable data, diverse goals) need autonomous workflows. Estimating variability from historical task data or domain expertise is the most reliable way to choose the right pattern.

Hybrid Workflow Patterns

Most production agents use hybrid patterns that combine elements from multiple workflow types. A common hybrid is a branching workflow with autonomous subtasks: the overall workflow follows a predetermined branching structure, but within each branch, the agent has autonomy to decide how to accomplish the specific subtask. This provides the predictability of branching for the overall structure with the flexibility of autonomy for the details.

Another common hybrid is a linear workflow with parallel data gathering. The overall workflow follows a fixed sequence (research, analyze, report), but the research phase uses parallel execution to gather data from multiple sources simultaneously. The analysis phase processes all gathered data sequentially because it requires the full dataset. The reporting phase might use parallel execution again to generate different sections of the report simultaneously. This selective use of parallelism optimizes performance without sacrificing the sequential logic where it matters.

Workflow composition builds complex workflows from simpler ones. A content creation workflow might compose a research workflow, a writing workflow, and a quality assurance workflow into a pipeline. Each component workflow can be developed, tested, and maintained independently, then combined into larger workflows as needed. This composability reduces development effort and improves reliability because each component is tested thoroughly before being used in composition.

Workflow selection can be dynamic, with the agent choosing the appropriate workflow pattern based on the task characteristics. A classification step at the beginning of the interaction determines whether the task is simple (single-step), routine (linear), variable (branching), or novel (autonomous). This meta-workflow selects the right execution pattern for each task, optimizing the tradeoff between efficiency and flexibility.