- In short
- Five factors decide between the augmented LLM, workflow, and agent patterns: predictability (how structured and repeatable the task is), error cost (how expensive a wrong output is), observability (how visible the work is while running), latency, and cost. More predictable tasks and higher error costs favour tighter control; less predictable tasks favour an agent, which typically takes more steps and costs more per request.
The factors that pick the pattern
Placing patterns on the autonomy spectrum tells you what the options are; the five deciding factors tell you which one to choose. The Claude Certified Architect - Professional (CCAR-P) exam treats weighing these factors as an analyse-level skill because a defensible pattern choice is grounded in them, not in preference or habit. Predictability, error cost, observability, latency, and cost each pull toward more or less autonomy, and reading how they apply to a specific problem is how you justify the augmented LLM, workflow, or agent you recommend.
Predictability is how structured and repeatable the task is: a highly structured task with a knowable sequence favours a fixed workflow, while an open-ended task whose steps cannot be predetermined favours an agent. Error cost is how expensive a wrong output is: higher error cost favours the tighter control of a workflow over the open-ended autonomy of an agent, because control is what limits how far a mistake can propagate. These first two often dominate.
- The five deciding factors
- The factors that choose between the augmented LLM, workflow, and agent patterns: predictability (how structured and repeatable the task is), error cost (how expensive a wrong output is), observability (how visible the work is while running), latency, and cost. More predictability and higher error cost favour tighter control; less predictability favours an agent.
Observability, latency, and cost
Observability is how visible the work is while it runs. A fixed workflow is easy to observe step by step, because the steps are known in advance and each can be logged and inspected. An open agent loop is harder to observe, because the sequence is decided at runtime and can vary run to run. When a problem demands that the work be auditable or debuggable as it happens, observability pulls toward a workflow.
Latency and cost are the operational factors. Agents typically take more steps and make more model calls per request than a fixed workflow tuned to the exact task, so they tend to cost more and add latency. A workflow that hits the task precisely does less work per request. When a tight latency budget or a hard cost ceiling is in play, these factors pull away from the agent end of the spectrum unless the task's unpredictability genuinely requires it.
Weigh, do not average
The critical discipline is that the five factors are not equally important for every problem. Treating them as a checklist to average is the mistake; the skill is identifying which factor is the binding constraint for this specific problem and letting it carry the most weight. A task with very high error cost may be settled by that factor alone, regardless of how the other four fall, because no amount of flexibility justifies an open loop on an action that must not go wrong. Weighing means finding the factor that dominates, not scoring all five and taking a mean.
What the CCAR-P exam trips candidates on
The exam tests two traps. The first is weighing all five factors as equally important instead of identifying which one is the binding constraint for this specific problem. A scenario will lay out all five and invite a balanced-sounding average, and the credited answer picks out the factor that dominates, often error cost or observability, and lets it decide.
The second is choosing an agent pattern purely because it is more flexible, without checking whether error cost or observability requirements rule it out. Flexibility is attractive, but a high-error-cost or high-observability requirement can disqualify an agent regardless of how flexible it would be. The reliable reading checks whether any factor rules the agent out before selecting it for its flexibility.
Worked example
A partner wants to automate the final approval of wire transfers above a large threshold. The task is fairly structured, but a wrong approval is extremely costly and hard to reverse, and auditors require that every decision be inspectable step by step. A designer proposes an agent 'because it can adapt to unusual cases.' Evaluate the choice against the five factors.
Resist averaging the factors into a balanced verdict. Two of the five, error cost and observability, are extreme here, and the skill is to let the binding one dominate rather than scoring all five equally. A wrong wire-transfer approval above a large threshold is enormously costly and hard to reverse, which is exactly the condition under which error cost favours the tightest control available, not an open-ended agent loop.
Observability compounds the point. Auditors require every decision to be inspectable step by step, and a fixed workflow is far easier to observe than an agent whose sequence is decided at runtime and varies run to run. Even the task's moderate structure argues against an agent, since a predictable, structured task is a good fit for a tuned workflow. The designer's justification, flexibility to adapt to unusual cases, is the exact trap: choosing an agent for its flexibility without checking whether error cost or observability rules it out. Here, both do.
The defensible choice is a tightly controlled workflow, or deterministic code for the approval decision itself, with any genuinely unusual cases escalated to a human rather than handled by an open agent loop. The binding constraint is error cost, reinforced by observability, and it settles the pattern regardless of the flexibility an agent would offer.
Common misreadings to avoid
Misconception
A good pattern decision weighs all five factors equally and picks the best average.
What's actually true
Misconception
An agent is the safe default because it is the most flexible pattern.
What's actually true
How this shows up on the exam
Domain 1 questions on this knowledge point describe a task with several competing pressures and ask which pattern fits. The reliable reading names all five factors, predictability, error cost, observability, latency, cost, but resists averaging them, instead identifying the factor that binds for this specific problem and choosing the pattern that respects it, especially checking whether error cost or observability rules out an agent.
These factors build on the autonomy spectrum and the workflow sub-patterns, and they sharpen into the tightest-constraint decision rule. Error cost is shared with automation vs augmentation, and latency and cost connect to cost as a hard design constraint.
A research assistant explores an open-ended question that no fixed sequence of steps can anticipate; outputs are low-stakes internal drafts, the work does not need step-by-step auditing, and there is generous latency and budget headroom. Which pattern fits and what makes it defensible?
People also ask
What are the five factors for choosing an AI pattern?
How does predictability affect pattern choice?
Why do agents cost more than workflows?
Watch and learn
Official Anthropic Academy lessons first, then hand-picked walkthroughs. Videos load only when you press play.
No videos curated for this concept yet
We are still curating the best official and community videos for this topic.
Official prep for this domain
Anthropic's own free prep module for this part of the syllabus, on the official prep course. Free with an Anthropic Academy sign-in.
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.