AnthropicClaude Certified Architect – Professional2026-101.5 Apply decomposition techniques for complex problem solving
Choose when findings may reshape the plan
Completed by the knowledge check
Learning objective
Choose fixed, predetermined decomposition or adaptive decomposition driven by intermediate findings.
Core idea
Keep known obligations in application-controlled stages; let findings determine subsequent investigations when you cannot specify them in advance. At production scale, defend that discretion through observable outcomes, bounded effort, and an owner for exceptions.
Anthropic distinguishes predetermined prompt chains from orchestrator-workers that determine subtasks from the input. Your decision is whether planning freedom improves the business result enough to justify its operating cost.
How it works
Establish the fixed obligations first
Write down what every accepted run must establish. In a software migration, compatibility evidence and passing tests remain required even when the files needing investigation differ. Anthropic’s prompting guidance identifies intermediate inspection and enforced pipeline structure as reasons for explicit chaining.
A known validation failure leading to a designed repair stage remains a fixed route. Grant adaptation where evidence creates necessary investigations, rather than merely selecting among predefined branches. Keep the assignment attached to its original decision through the objective, output format, source guidance, and boundaries described in Anthropic’s research account. See 1.5.1 for independently verifiable completion.
Price the planning freedom
Anthropic’s research-system account reports that, in its data, agents typically used about 4× more tokens than chat interactions, and multi-agent systems about 15× more than chats. These are observations from that system, not multipliers to apply as a pricing formula. The post argues that tasks must be valuable enough to pay for the increased performance.
It also identifies dependency-heavy work as a weaker fit: most coding tasks have fewer truly parallelizable parts than research. A migration may need adaptive investigation without needing many concurrent workers. Evaluate whether separate assignments add useful coverage or repeatedly exchange the same shared context.
Bound the Agent SDK investigation
For a planner running on the Agent SDK, set max_turns in Python or maxTurns in TypeScript to limit tool-use turns. Set max_budget_usd in Python or maxBudgetUsd in TypeScript for an estimated-spend threshold. Neither option is a wall-clock deadline; the application must enforce its delivery deadline separately.
Hitting a limit returns a ResultMessage with subtype error_max_turns or error_max_budget_usd. Route those outcomes to the incomplete-investigation path. They do not carry a successful final result; preserve accepted evidence separately so the reviewer can inspect what remains unresolved.
Leave budget headroom. Spend is checked after model responses arrive, so the cap can be exceeded by the response that reaches it and by running subagents’ spend before they stop. Subagent spend counts toward the total; enforcement that refuses additional spawns and stops background subagents requires Claude Code v2.1.217 or later, as documented by the agent-loop page. A spend cap is an operating control, not an exact prepaid ceiling.
Record the choice your team can defend
Record the uncertainty requiring adaptation, retained fixed stages, and evaluation evidence. Anthropic’s research account evaluates outcomes because different paths can validly reach the same goal. Compare designs on routine cases, consequential discoveries, and investigations that cannot settle the question.
Record accepted outcomes, missing requirements, elapsed time, and consumed work. Assign the workflow owner to fixed gates, the planner team to assignment prompts and evaluation cases, and a domain owner to unresolved judgments. Those owners must be able to change the design when its cost or coverage fails acceptance.
Key distinctions and trade-offs
The choices have different operational consequences:
| Choice | Fits the requirement | Cost to defend |
|---|---|---|
| Fixed | Required stages and handoffs are known before execution | Extra calls and downstream dependence on earlier outputs |
| Adaptive | Intermediate evidence reveals necessary investigations | Variable paths and effort; broader evaluation of valid outcomes |
| Hybrid | Acceptance stages are stable but discovery is variable | An explicit interface between exploratory findings and fixed gates |
Keep acceptance criteria stable even when investigations vary. More delegation is justified by measured value, not by the presence of an adaptive planner.
Architecture / decision scenario
A technology company operates a migration assistant for six product teams. Each change must reach review within one business day and include compatibility evidence and test results. Most changes touch known adapters; a minority reveal additional consumers during inspection.
The architect keeps compatibility and testing as fixed stages and permits adaptive consumer discovery. Dependency-heavy code analysis stays with one investigator unless a separate assignment has complete inputs. The team compares this hybrid with a fixed-only baseline on routine and hidden-consumer cases, measuring coverage and token consumption.
For this workload, the team selects illustrative TypeScript limits of maxTurns: 30 and maxBudgetUsd: 5, with budget headroom and a separate application deadline. An error_max_turns or error_max_budget_usd result sends accepted findings and unresolved consumers to review; it cannot pass the compatibility gate. The one-business-day deadline uses the same incomplete-delivery path.
The migration platform team owns planner changes and gates; product maintainers accept unresolved compatibility decisions. The hybrid is retained only if additional coverage justifies its measured cost within the delivery constraint.
Common exam traps
- Choosing adaptation because a task is long, although required stages are known.
- Treating a predetermined conditional branch as autonomous planning.
- Assuming coding work always benefits from more parallel agents.
- Treating a spend threshold as an exact ceiling or wall-clock deadline.
- Calling a limit subtype successful completion.
What to remember
Keep known acceptance obligations fixed. Grant adaptive decomposition when findings create necessary work, and defend it through measured coverage, cost, and delivery time. Name SDK turn and spend limits and handle their error subtypes as incomplete outcomes.
Knowledge check
An accounts-payable workflow runs every invoice through extraction, validation, and posting. Validation reports only known failure types, and each one routes to a designed repair stage before posting. A team proposes replacing the chain with an orchestrator because some invoices take the repair route. What should the architect decide?
A migration planner runs on the TypeScript Agent SDK. Finance sets a planning target of $5 per run, allowing for documented SDK overshoot, and the product team requires the report within two hours. Which configuration addresses both concerns?
maxBudgetUsd: 5 and treat it as the run’s deadline as wellmaxBudgetUsd: 5 and size maxTurns to enforce the deadlinemaxTurns , sized to keep the run within both requirementsmaxBudgetUsd below 5 and keep the deadline in application codemaxBudgetUsd nor maxTurns is a wall-clock limit: maxTurns counts tool-use turns, so the application must enforce the delivery deadline itself.An Agent SDK investigation of hidden consumers ends with a ResultMessage whose subtype is error_max_turns. Before stopping, it accepted evidence for three consumers, and the full number is unknown. What should the pipeline do with this run?
result text to the compatibility gateresult and cannot pass a gate that requires complete compatibility evidence. Evidence accepted along the way is preserved separately and sent to the incomplete-investigation path, where a reviewer sees what remains unresolved. Repeating the run unchanged discards that evidence and is likely to stop at the same limit.A code-migration assistant must discover previously unknown consumers and splits each change across eight parallel subagents, one per module. Logs show the modules share a call graph: subagents repeatedly request the same shared context, and coverage is no better than a single investigator’s. Which change should the architect evaluate first?
References
- Anthropic, Building effective agents
- Anthropic, Prompting best practices
- Anthropic, How we built our multi-agent research system
- Anthropic, How the agent loop works