Task 1.2: Orchestrate multi-agent systems with coordinator-subagent patterns

• AHave both subagents append to a shared scratch file that synthesis reads when it starts

  Incorrect

  A shared file lets data bypass the coordinator, so failures and partial writes are no longer seen or handled in one place.

• BKeep every exchange going through the coordinator, which decides what to pass on and handles failures

  Correct

  Routing everything through the coordinator gives one point for observability, error handling and control of the information flow, which is what the requirements ask for.

• CMerge document analysis and synthesis into a single subagent that holds both tool sets, so no hand-off is needed at all

  Incorrect

  Merging removes the hop but also the separation of concerns, and the merged agent carries more tools and a larger context than either job needs.

• DExpose each subagent to the others as a tool on a shared MCP server so they can call each other when they need something

  Incorrect

  Peer-to-peer calls spread routing and error handling across agents, so no single place sees or controls the flow.

A hub-and-spoke design keeps the coordinator in charge of decomposition, delegation, error handling and aggregation. Direct links between subagents trade that control and traceability for one saved hop.

• AGive each subagent a distinct slice of the topic or source types, with its boundaries stated in the delegation

  Correct

  Distinct subtopics or source types with explicit boundaries stop subagents from repeating each other and make gaps visible before the work starts.

• BTell each subagent to be thorough and to avoid results the other subagents are likely to find

  Incorrect

  A subagent cannot see its siblings' work, so it has no way to know what to avoid unless the coordinator divides the scope.

• CKeep the same instruction but add a deduplication step to the synthesis subagent before it writes the report

  Incorrect

  Deduplication hides the overlap after the tokens are spent and does nothing for the source type no one searched.

• DSpawn five subagents with the same instruction instead of three, so that more of the topic is covered by the extra searches

  Incorrect

  More agents with the same vague task mostly produce more of the same results and raise the token bill further.

Partition the research scope across subagents: give each a different subtopic or source type and clear boundaries. Vague, identical delegations lead to duplicated searches and missed areas.

• AThe agents' decision patterns and interaction structure: which subagents ran, which tools they called and how results flowed, without conversation content

  Correct

  Anthropic traced decision patterns and interaction structures, without monitoring the contents of individual conversations, to find root causes while protecting privacy.

• BFull transcripts with user names masked, so engineers can read exactly what each agent saw

  Incorrect

  Masking names still exposes the conversations' content, which the policy forbids.

• COnly the final reports, since they show what the system produced

  Incorrect

  The final report shows the symptom; it cannot tell you whether queries, sources or tools caused it.

• DNothing new; rely on users' thumbs-down ratings to find the bad runs

  Incorrect

  Ratings point to bad runs but not to the coordination step that failed.

Routing work through a coordinator makes the system observable. Trace the structure of each run, the delegation decisions, tool calls and hand-offs, to diagnose coordination failures; this does not require reading conversation content.

• ASpawn subagents immediately and let each one report back how many more subagents it thinks are needed

  Incorrect

  This adds more delegation without planning; the overlaps and wrong tool choices come from the coordinator never working out the plan first.

• BHard-code the plan: always three subagents with the same roles and tools for every query

  Incorrect

  A fixed plan removes the misfires on some queries and creates them on others, since research queries differ in what they need.

• CHave the coordinator plan in extended thinking first: which tools fit, how complex the query is, and each subagent's role

  Correct

  Anthropic's lead agent uses thinking to assess tool fit, judge query complexity and subagent count, and define each subagent's role before delegating.

• DLet the synthesis subagent decide the plan after it sees the first round of results

  Incorrect

  Planning after the first round still spends that round on misassigned subagents, and synthesis is not the role that delegates.

Coordinators decompose and delegate better when they plan before acting. Extended thinking gives the lead agent a scratchpad to choose tools, judge complexity and the number of subagents, and define distinct roles, so delegation is deliberate rather than corrected mid-run.

• AHave the synthesis agent rank findings by how many sources repeat them

  Incorrect

  Content farms copy each other, so counting repetitions rewards the same low-quality pages.

• BRestrict the subagents to an allowlist of five trusted domains

  Incorrect

  An allowlist cuts off the open web and the authoritative sources that are not on it.

• CTell the subagents to use longer, more specific search queries

  Incorrect

  Overly long, specific queries tend to return few results, and they say nothing about which sources to trust.

• DAdd source-quality heuristics to the subagents' prompts

  Correct

  Explicit guidance on what makes a source credible counters the bias toward highly ranked pages while keeping the web in scope.

Anthropic's testers saw early research agents choose SEO-optimized content farms over authoritative but lower-ranked sources; adding source-quality heuristics to the prompts fixed it. Subagents need to be told what a good source looks like.

• AName the same output fields in every delegation, such as target, deadline and funding

  Correct

  Anthropic's research system gives each subagent an output format along with its objective; when every result arrives with the same fields, the coordinator can combine them.

• BAdd more background on EV charging policy to each delegation so that the subagents understand the topic better

  Incorrect

  Background helps a subagent with its objective, not with what to return, so the results would still come back in different shapes.

• CHave the coordinator convert each result into the table's format itself before it compares them

  Incorrect

  Reformatting cannot supply figures a subagent never reported, and it spends the coordinator's context on cleanup.

• DGive all the countries to a single subagent so that one agent reports them the same way

  Incorrect

  This gives up the parallel searches and separate context windows that suit a breadth-first comparison, when stating the format fixes the problem directly.

Aggregation starts with delegation. Anthropic's account of its research system names an output format among the things each subagent's task description needs. When the coordinator has to combine results, a shared format in every delegation makes them comparable.

• AAsynchronous execution removes the need for the coordinator to route results between subagents

  Incorrect

  The coordinator still has to collect and route results; asynchrony makes that coordination harder, not unnecessary.

• BSynchronous execution lets the coordinator steer subagents while they are still searching

  Incorrect

  In the synchronous setup the lead agent cannot steer subagents mid-run; it waits for them to finish.

• CAsynchronous execution always costs less because subagents finish sooner

  Incorrect

  Concurrency changes timing, not the work done, and it adds coordination overhead.

• DSynchronous execution simplifies coordination, but one slow subagent can block the whole system

  Correct

  Waiting for each set keeps coordination simple, at the price of bottlenecks while a single subagent finishes.

• EAsynchronous execution adds challenges in result coordination, state consistency and error propagation across subagents

  Correct

  More parallelism means the coordinator must reconcile results, keep state consistent and handle errors arriving at any time.

In a hub-and-spoke system the coordinator owns communication, routing and error handling. Synchronous execution keeps that simple but creates bottlenecks; asynchronous execution adds parallelism and makes result coordination, state consistency and error propagation harder.