- In short
- Choosing a task decomposition strategy means matching the approach to the task: a fixed sequential pipeline for predictable, structured work; dynamic adaptive decomposition for open-ended investigation with unknown scope; and a multi-pass split when input volume threatens consistent quality.
What choosing a decomposition strategy involves
This is the evaluate-level knowledge point of task statement 1.6, and it sits on top of the three building blocks: fixed sequential pipelines, dynamic adaptive decomposition, and the multi-pass remedy for attention dilution. Choosing well is not about knowing which strategy is "best" in the abstract, none is. It is about reading the characteristics of the task in front of you and matching the approach to them, then being able to justify the match.
A good task decomposition strategy decision turns on a few diagnostic questions. Is the work predictable, with stages you can name in advance, or is its scope unknown until you start? Does the input volume threaten consistent quality? How much do you value adaptability versus predictable cost and auditability? Each answer pushes you toward a different approach, and the exam wants you to make that reasoning explicit.
- Choosing a decomposition strategy
- The evaluative skill of selecting fixed, dynamic, or multi-pass decomposition by reading a task's predictability, scope, input volume, and adaptability requirements, then justifying the choice against those characteristics rather than applying a fixed default.
The three strategies and the signals that pick them
Each strategy has a characteristic profile. Learn the signals that point to each, and most exam scenarios resolve quickly.
Fixed sequential pipeline is for predictable, structured tasks where the stages are known and consistent output matters. Signals: a repeatable process, well-defined inputs and outputs, a need for auditability and predictable cost. Anthropic frames this as the workflow end of the spectrum, fixed subtasks along predefined paths.
Dynamic adaptive decomposition is for open-ended investigation where scope is unknown and the system must adapt. Signals: "the cause is unknown," "explore," "the structure isn't documented," a need to follow discoveries. This is the agent end of the spectrum, where the model directs its own process.
Multi-pass split is for large input volume where quality consistency is critical. Signals: many files or records to analyse, a risk of inconsistent depth, contradictory verdicts on equivalent items. The split, focused per-item passes plus a cross-item integration pass, is the structural answer to attention dilution.
A decision flow you can apply under pressure
The fastest way to choose is to ask the questions in order of discriminating power. First, is the scope known? If the steps cannot be enumerated in advance, you are in dynamic-decomposition territory regardless of anything else. If the scope is known, ask next whether input volume threatens consistent depth; if it does, a multi-pass split is required even though the task is otherwise predictable. If neither condition holds, known scope, manageable volume, a fixed pipeline is the simplest tool that works, and simplest-that-works is the right instinct.
Anthropic's own guidance reinforces this ordering: prefer the simplest approach that meets the need, and only add the autonomy of an agent when predefined steps cannot capture the task. Reaching for a complex strategy when a simple one suffices is as much a mistake as the reverse.
Combining strategies within one workflow
The exam's framing of "choosing" can make it sound like you pick exactly one strategy per task and stop. In real workflows the strategies nest, and the evaluate level rewards architects who can see that. A large, open-ended task often uses dynamic adaptive decomposition at the top level to plan the investigation, while individual subtasks inside it are executed with a fixed pipeline or a multi-pass split. The question is rarely "which one strategy," but "which strategy at which level."
Consider a research agent asked to assess an unfamiliar service for production-readiness. The top level is genuinely open-ended, you cannot enumerate what you will find, so the agent decomposes dynamically, generating probes from discoveries. But when one of those probes turns out to be "audit these 40 configuration files for the same set of risks," that subtask is predictable and high-volume, so it is best run as a multi-pass split rather than another round of exploration. And a final "compile the findings into the standard report" step is a fixed pipeline because its stages never change. One workflow, three strategies, each matched to the characteristics of its level.
Seeing combinations also guards against a subtler default: treating the whole task as homogeneous. A task is rarely uniformly predictable or uniformly open-ended. The skill is to decompose far enough to notice that different parts have different characteristics, then apply the matching strategy to each part. That is why the prerequisites for this knowledge point include all three building blocks. You cannot choose well among options you do not individually understand, and you certainly cannot combine them.
A practical tell that you are facing a combination rather than a single strategy is the word "then" in the task description: "investigate, then fix," "map, then refactor," "gather, then summarise." Each "then" marks a seam where the characteristics may shift, the first half open-ended, the second half structured, and where a different strategy may belong. Training yourself to hear those seams is a fast route to the right decomposition under exam time pressure.
Underneath every choice sits one principle from Anthropic's guidance: prefer the simplest approach that meets the need, and add complexity only when the task demands it. A fixed pipeline is simpler than a dynamic agent; a single well-structured pass is simpler than a multi-pass split. Reach for the more elaborate strategy when, and only when, a characteristic of the task forces it. Justifying a choice, at the evaluate level, means naming both the characteristic that forced the complexity and the simpler option you rejected.
Why a single default is the classic mistake
The exam trap for this knowledge point is defaulting to one strategy for all tasks without considering characteristics. Architects who love agents reach for dynamic decomposition on a tidy, repeatable job and pay for it in unpredictable cost and harder auditing. Architects who love pipelines force a fixed chain onto an open-ended investigation and watch it sail past the real problem. Both are the same error: substituting a personal preference for an evaluation of the task.
Evaluating properly also means recognising when characteristics combine. A large, open-ended investigation might use dynamic decomposition to plan the work and a multi-pass split to execute a high-volume analysis subtask within it. The strategies are not mutually exclusive; they are tools you select per task and sometimes per sub-task. Being able to defend a combination against the task's characteristics is the mark of the evaluate level.
Worked example
An architect is handed three tasks in one sprint and must justify a decomposition strategy for each: (1) a nightly invoice-formatting job, (2) diagnosing a sporadic outage of unknown cause, (3) a security audit of a 40-file service.
The skill being tested is matching strategy to characteristics and justifying each choice, so the architect evaluates them one at a time.
Task one, the nightly invoice job, is predictable and structured: the same stages run every night and consistency plus auditability are the priorities. The scope is fully known and the volume is modest, so a fixed sequential pipeline is correct. Choosing dynamic decomposition here would add variability for no benefit.
Task two, the sporadic outage, has unknown scope, the cause is the very thing to be found. No fixed chain can branch toward an undiscovered fault, so dynamic adaptive decomposition is correct. The agent generates each next probe from what the last one revealed.
Task three, the 40-file audit, has known scope (check each file for known vulnerability classes) but a volume large enough to threaten consistent depth, the attention-dilution risk. The right choice is a multi-pass split: focused per-file passes for uniform local depth, then a cross-file integration pass for issues that span files. Note that task three is predictable yet still not a plain fixed pipeline, because volume, not unpredictability, is the deciding characteristic. Articulating exactly which characteristic drove each decision is what an evaluate-level answer requires.
To close, the architect is handed a fourth, messier task: investigate why a data export occasionally produces corrupt files, then harden whatever is found. This one resists a single label, and that is the lesson. The investigation half has unknown scope, so it is dynamic; if the cause turns out to be a validation gap spread across many export handlers, hardening them is a high-volume, consistency-critical job better run as a multi-pass split; and emitting the final remediation report is a fixed pipeline. The architect's answer is not one strategy but a justified composition, dynamic at the top, multi-pass and fixed underneath, with each choice tied to the characteristic that drove it. The word "then" in the brief was the giveaway that two halves with different characteristics were hiding inside one sentence. That layered justification is exactly what an evaluate-level question is probing for.
Common misconceptions
Misconception
Dynamic decomposition is the most advanced strategy, so a strong architect should default to it for complex tasks.
What's actually true
Misconception
If a task is predictable and structured, a fixed pipeline is always the right choice.
What's actually true
How this shows up on the exam
Expect scenario stems that describe a task in enough detail to reveal its characteristics, then ask which decomposition strategy to use and why. Because the Bloom level is evaluate, the credited answer is the one whose justification matches the signals: unknown scope earns dynamic decomposition, manageable predictable work earns a fixed pipeline, and high-volume consistency-critical work earns a multi-pass split. Distractors usually pick a defensible-sounding strategy with the wrong justification, or apply a single default to a task whose characteristics call for something else. Train yourself to name the deciding characteristic out loud, scope, then volume, then everything else, and to defend combinations when a task has more than one. That discipline connects directly to broader context-management strategy selection when the workflow grows large.
A team must build an agent that reviews a 50-file pull request for a fixed checklist of code-quality issues. The checks never change, but reviewers complain the current single-pass agent catches an issue in some files and misses the same issue in others. Which decomposition strategy is the best justified choice?
People also ask
How do you choose a task decomposition strategy?
When should you use a workflow instead of an agent?
What task characteristics decide fixed versus dynamic decomposition?
Watch and learn
Official Anthropic Academy lessons first, then hand-picked walkthroughs. Videos load only when you press play.
Agents and workflows
Why watch: Helps you choose between a flexible agent and a fixed pipeline based on task predictability, the decomposition-strategy choice.
More videos for this concept
References & primary sources
Master this concept with Archie
Practice it inside an adaptive study session. Archie, your Socratic AI tutor, tracks your mastery with Bayesian Knowledge Tracing and schedules the perfect next review.