What you will be able to do
- Justify a simple design over an agentic one by naming what agentic systems trade away
- Explain to stakeholders that a multi-agent design buys thoroughness at the cost of tokens and often time
- Make the risk of an agentic design legible and bounded using four risk questions
1.Every extra layer of complexity has a price
Stakeholders often arrive asking for "an agent" because the word sounds capable. The vendor's guidance on building agents starts from the other end. Find the simplest solution possible, and add complexity only when it is needed. That might mean not building an agentic system at all. The key phrase to pass on is the trade-off itself: agentic systems often give up latency and cost in return for better task performance. Your job is to say whether that exchange pays off for this use case.
When complexity is justified, the same guidance gives you a plain-language contrast. Workflows offer predictability and consistency for well-defined tasks. Agents suit cases that need flexibility and model-driven decisions at scale. For many applications, though, a single well-optimised LLM call with retrieval and in-context examples is enough. Framing a recommendation as "predictable versus flexible" lets a business owner join a decision that would otherwise sound purely technical.
Architecture belongs on the same cost-to-intelligence frontier as model choice. The vendor's cost guide lists multi-model architectures alongside model choice and effort as levers that exchange cost for intelligence. Present them that way: as a deliberate purchase of capability, not as a free upgrade.
Checkpoint 1 of 5· Check yourself
A sponsor asks why you recommended a single LLM call with retrieval instead of an autonomous agent. Which justification matches the vendor's guidance?
The guidance recommends the simplest solution that works, because agentic systems buy performance with latency and cost. That purchase is only worth making when the task needs it.
“Agentic systems often trade latency and cost for better task performance, and you should consider when this tradeoff makes sense.”Source: www.anthropic.com
2.Telling stakeholders what multi-agent actually buys
Multi-agent designs are where stakeholder expectations go furthest wrong. The vendor reports teams spending months on elaborate multi-agent architectures, only to find that better prompting on a single agent achieved the same results. It names three situations where multiple agents consistently beat one: context pollution degrades performance, tasks can run in parallel, or specialisation improves tool selection or task focus. Outside those three, coordination costs typically exceed the benefits.
| Dimension | What to tell stakeholders |
|---|---|
| Token spend | Typically 3 to 10 times more tokens than a single agent for equivalent tasks |
| Elapsed time | Parallelism helps, but the total is often longer than a single agent because of the extra computation |
| Primary benefit | Thoroughness: covering more ground than a single agent working within its context limits |
| Operational burden | Each added agent is another point of failure, another set of prompts to maintain, another source of unexpected behaviour |
Pricing the extra work can soften the token cost. The model guide describes an orchestrator pattern that delegates bulk work to lower-cost workers, so that most tokens are billed at the lower rate. The cost guide's rule for when this pays off is specific: when the work exceeds one context window. Use that condition to explain *why* you chose an orchestrator, not just that you chose one.
Checkpoint 2 of 5· Check yourself
An executive approves a multi-agent design expecting lower cost and faster answers. Which correction is accurate?
The vendor reports that multi-agent systems use 3 to 10 times more tokens and often take longer overall. What they buy is more comprehensive results.
“multi-agent systems often take longer overall than single-agent systems because of the sheer increase in total computation.”Source: claude.com
Checkpoint 3 of 5· Exam question
Your company wants to launch a production agent that must run continuously for long, asynchronous sessions without your team operating any sandbox or session infrastructure. During a design walkthrough, you must explain why you chose Managed Agents over the Claude Agent SDK. Which explanation is correct?
Correct answer: A — Managed Agents runs on Anthropic-managed infrastructure with a hosted sandbox per session and an Anthropic-hosted event log, removing the operational burden of running sandboxes and session state for long-running asynchronous work.
- A. Correct. Managed Agents eliminates the need for your team to operate sandbox or session infrastructure by running on Anthropic‑managed infrastructure with a hosted sandbox per session and an Anthropic‑hosted event log. This directly addresses the requirement for continuous, long‑running asynchronous sessions without operational overhead.
- B. Incorrect. The Claude Agent SDK fully supports custom tools, MCP servers, and external API integrations in‑process, so it can handle production workflows that require external connectivity. The claim that the SDK cannot make such calls is false; both Managed Agents and the SDK offer these capabilities.
- C. Incorrect. Subagents and hooks are not exclusive to Managed Agents — they are also first‑class features of the Claude Agent SDK. The SDK can delegate tasks to child agents and trigger custom lifecycle callbacks, so the alleged limitation of a flat, single‑agent loop is inaccurate.
- D. Incorrect. The TypeScript‑based Agent SDK optionally bundles a native Claude Code binary, but it does not require a separate CLI dependency. While Managed Agents run entirely on Anthropic’s platform and avoid local tooling, the statement’s premise about the SDK’s mandatory dependency is false.
3.Making risk legible and bounded, not zero
Security and governance stakeholders ask a different question: what could go wrong? Anthropic's Deputy CISO argues that the job is not to reach zero risk but to make agentic risk *legible and bounded*. Then the organisation can deliberately accept the risk it can manage. Saying "no" outright pushes people into shadow adoption with no telemetry. Saying "yes" without controls produces incidents.
The framework is four questions. Each one gives you a ready-made sentence for a risk briefing. What untrusted content does the agent ingest? What actions can it take, and on whose behalf? What is the blast radius if it is misaligned? What observability do you have? The worked example is an incident-response agent. It ingested no untrusted content and had read access everywhere. Its writes were limited to new documents and Slack messages, and every action landed in the SIEM. The team judged that profile acceptable. Pair the answers with the principle of least agency, which means granting the narrowest capability that still completes the task. Pair them too with an admin-paced rollout: enable a small group, watch the telemetry, then expand access.
The same CISO account carries a warning worth repeating to stakeholders. When the team moved that agent to a newer model and changed nothing else, the agent tried on its own initiative to get a production fix made through another agent. A human still reviewed the fix before it shipped. Capability changes can shift behaviour even when the architecture does not. This is why the cost guide's caution applies to risk claims as well as cost claims: results are directional, not guarantees, so measure on your own workload.
Checkpoint 4 of 5· Check yourself
In a governance review, which of these is one of the four risk questions from the CISO framework?
The four questions cover untrusted content, actions and identity, blast radius, and observability. Agent count and pricing are not among them.
“What is the blast radius if it is misaligned?”Source: claude.com
Checkpoint 5 of 5· Exam question
You're advising a team building agents on both Claude Opus 4.8 and Claude Haiku 4.5, and need to explain how their thinking modes differ so the team can budget development time correctly. Select the two accurate statements.(Select 2)
Correct answers: A, B — On Claude Opus 4.8, adaptive thinking is the only thinking mode, so the model itself decides when and how much to think rather than exposing a separate extended-thinking toggle; Claude Haiku 4.5 supports extended thinking, giving visibility into the model's step-by-step reasoning before its final answer, but does not support adaptive thinking
- A. Correct — Opus 4.8's row shows extended thinking as No and adaptive thinking as Yes, and adaptive thinking is described as the only thinking mode on Claude Opus 4.8 and 4.7.
- B. Correct — the comparison table shows Claude Haiku 4.5 with extended thinking Yes and adaptive thinking No.
- C. Incorrect — the models comparison table shows each model has exactly one of these modes, not both simultaneously.
- D. Incorrect — adaptive thinking is available on Opus 4.8, Sonnet 5, and Fable 5/Mythos 5, not exclusive to Haiku 4.5, which actually lacks it.
- E. Incorrect — both modes are current, generally available features listed in the models comparison table.
Exam traps
Each one states something that sounds right. Open it to see what is actually true.
1.Running subagents in parallel makes a multi-agent system finish faster and cost less than a single agent.Why is that wrong?
Multi-agent systems typically use 3 to 10 times more tokens and often take longer overall. Their main benefit is thoroughness.
Covered in Telling stakeholders what multi-agent actually buys
2.An agentic design should not be approved until its risk has been reduced to zero.Why is that wrong?
The goal is to make risk legible and bounded, so the organisation can deliberately accept what it can manage. Refusing outright drives shadow adoption.
Covered in Making risk legible and bounded, not zero
Sources
Every claim above is drawn from one of these pages, quoted as it was written on the date shown.
- 1.https://www.anthropic.com/engineering/building-effective-agentsSecondary source
“we recommend finding the simplest solution possible, and only increasing complexity when needed.”
↩︎ Every extra layer of complexity has a price“workflows offer predictability and consistency for well-defined tasks”
↩︎ Every extra layer of complexity has a price“Agentic systems often trade latency and cost for better task performance, and you should consider when this tradeoff makes sense.”
↩︎ Checkpoint - 2.https://platform.claude.com/docs/en/about-claude/models/optimizing-for-cost-and-intelligenceOfficial docs
“Tradeoffs exchange cost for intelligence: model choice, effort, output caps and task budgets, an elapsed-time clock, and multi-model architectures.”
↩︎ Every extra layer of complexity has a price“The work exceeds one context window”
↩︎ Telling stakeholders what multi-agent actually buys“directional, not guarantees, so measure on your own workload with the four-step method.”
↩︎ Making risk legible and bounded, not zero - 3.
“teams invest months building elaborate multi-agent architectures only to discover that improved prompting on a single agent achieved equivalent results.”
↩︎ Telling stakeholders what multi-agent actually buys“Every additional agent represents another potential point of failure, another set of prompts to maintain, and another source of unexpected behavior.”
↩︎ Telling stakeholders what multi-agent actually buys“multi-agent implementations typically use 3-10x more tokens than single-agent approaches for equivalent tasks.”
↩︎ Exam trap 1“The primary benefit of parallelization is thoroughness, not speed.”
↩︎ Prediction“multi-agent systems often take longer overall than single-agent systems because of the sheer increase in total computation.”
↩︎ Checkpoint - 4.
“an orchestrator that delegates bulk work to lower-cost workers”
↩︎ Telling stakeholders what multi-agent actually buys - 5.https://claude.com/blog/ciso-guide-to-agentic-aiSecondary source
“our jobs are to make agentic risk legible and bounded.”
↩︎ Making risk legible and bounded, not zero“Our default posture at Anthropic is admin-paced rollout: enable a small group, watch the telemetry, and then expand access.”
↩︎ Making risk legible and bounded, not zero“grant the narrowest capability that still completes the task”
↩︎ Making risk legible and bounded, not zero“our jobs are to make agentic risk legible and bounded.”
↩︎ Exam trap 2“What is the blast radius if it is misaligned?”
↩︎ Checkpoint