AnthropicClaude Certified Architect – Professional2026-100 Prerequisites
Tell augmented calls, workflows, loops, and delegation apart
Completed by the knowledge check
Learning objective
Distinguish single augmented LLM calls, deterministic workflows, autonomous agent loops, and multi-agent delegation.
Core idea
A harness can implement an agent loop, but the architectural pattern depends on who selects the next step. This prerequisite adds the decision you defend to stakeholders: whether the task needs one enriched call, a prescribed path, model-directed iteration, or delegation across agents.
Anthropic recommends the simplest solution that meets the task. Extra autonomy earns its place when flexible decisions improve outcomes enough to justify added latency, cost, and opportunities for compounding errors.
How it works
The patterns answer different control questions:
| Pattern | Who selects the path? | What you must justify |
|---|---|---|
| Single augmented call | One bounded call; the model may choose a query or tool without directing an ongoing action loop. | Why that evidence is sufficient without model-directed follow-up. |
| Deterministic workflow | Code prescribes stages, routes, and gates around model calls. | Why the task structure is predictable enough to encode. |
| Autonomous agent loop | The model selects successive actions using environmental feedback. | Why the needed steps cannot be prescribed in advance. |
| Multi-agent delegation | A coordinator assigns subtasks and integrates worker results. | Why parallelism, specialization, or context isolation exceeds coordination cost. |
An augmented LLM can have retrieval, tools, and memory. The single-call pattern bounds use of these augmentations rather than giving the model ongoing control of execution; a model-selected query or tool does not alone imply an autonomous loop. See 1.3.1 for classification and 0.4 for those external connections.
“Deterministic workflow” describes control structure, not guaranteed identical model outputs. Anthropic’s workflow patterns include prompt chaining, routing, parallelization, orchestrator-workers, and evaluator-optimizer. A workflow can use model judgment inside a prescribed orchestration structure; orchestrator-workers even determines subtasks dynamically. See 1.3.2 for choosing among those patterns.
An autonomous loop instead uses tool results to decide what to do next and whether it has enough information. It needs explicit stopping conditions and checkpoints. See 1.3.3 for bounded autonomy and 1.3.4 for comparing blast radius, auditability, latency, and fallback behavior. The existence of a capable harness does not establish that a task needs this control pattern.
Delegation is another dimension: multiple workers can sit inside a workflow or an agent-driven process. In the Agent SDK, programmatic subagents use the agents option and are invoked through the Agent tool. Their intermediate work stays in separate conversations, and a final message returns to the parent. That isolates context but also makes the handoff important: the parent must receive enough evidence to integrate the result. See 1.4.1 for the multi-agent decision and 1.4.3 for delegation contracts.
Key distinctions and trade-offs
More calls are not necessarily more autonomy. A prescribed five-stage pipeline remains a workflow, while a short loop can make model-directed choices. More agents are not necessarily faster: independent subtasks can run concurrently, but dependent work still waits and integration adds work. Make the control path and dependencies explicit before comparing patterns.
Architecture / decision scenario
A retailer has a two-second target for routine return explanations and a separate queue for investigating unusual supplier discrepancies.
For routine answers, you retrieve the applicable policy before one augmented call. A fixed workflow validates the transaction and routes exceptions. You reserve an agent loop for discrepancies whose next lookup depends on previous findings. Multiple workers earn a pilot only when supplier investigations split into independent evidence searches; the coordinator must still reconcile their findings. You compare all four candidates against the same task outcomes under 1.3.4 rather than selecting delegation because it sounds more capable.
Common exam traps
Avoid these classification shortcuts:
- Equating tools or multiple calls with autonomous execution.
- Equating deterministic control flow with deterministic text.
- Assuming delegation removes the need to integrate and verify results.
What to remember
Classify control ownership first, then defend the added work. Augmentations provide capabilities; workflows prescribe paths; loops adapt actions; delegation distributes subtasks and creates a handoff obligation.
Knowledge check
A claims pipeline runs five fixed stages: classify the claim, extract fields, validate them in code, draft a letter, and route by a coded rule. Each model stage calls Claude, and the code decides every next step. Which pattern is this?
An IT operations team investigates unusual outages. Each diagnostic lookup depends on what the previous one found, and the steps cannot be listed in advance. Which design is the simplest that meets this task?
A parent agent built with the Agent SDK delegates supplier investigations to subagents defined in its agents option. Reviewers find the parent cannot verify the subagents’ findings before writing its report. Which change lets it verify them without repeating the searches?
References
- Anthropic, Building effective agents
- Anthropic, Subagents in the SDK