Task statement 1.2 · Agentic architecture and orchestration

Orchestrate multi‑agent systems with coordinator‑subagent patterns

In the research system, the coordinator decides which subagents run and what each one is asked. When that decision is wrong, every subagent can succeed and the report still silently skips part of the topic.

Why it's tested

When a report misses whole subtopics, the tempting fix is to broaden the queries of the subagent whose output looks thin. That fails when every subagent's output is correct, because the missing subtopics were never assigned to anyone. The same trust in the pipeline sends a one-fact question through all four subagents, which costs minutes and adds no coverage.

KnowledgeHub and spoke

Hub-and-spoke is an architecture in which the coordinator is the hub and each subagent sits on a spoke. The coordinator manages all communication, error handling and information routing, so search results reach synthesis only because it passes them on. Each subagent works in its own context window and does not inherit the coordinator's conversation history, so it knows only its prompt. Routing everything through the coordinator gives observability, because one log shows every delegation and result. It also gives consistent error handling and controlled information flow, because one agent decides what to retry and what each subagent sees.

KnowledgeWhat the coordinator decides, in order

The coordinator's role covers task decomposition, delegation, result aggregation, and deciding which subagents to invoke based on query complexity. Decomposition is its first split of the query into subtasks. Subagents cannot change it, because none of them sees the whole query. Only the coordinator can revise it, by reading the combined results for gaps and re-delegating, a loop called iterative refinement.

  1. Judge the query's complexity and select which subagent types it needs.
  2. Decompose it into subtasks, one distinct subtopic or source type each.
  3. Delegate each subtask with a Task tool call whose prompt is the full brief.
  4. Each subagent works in its own context window and returns its final message.
  5. Aggregate the results, invoke synthesis, and evaluate its output for gaps.
  6. Re-delegate targeted queries and re-invoke synthesis until coverage is sufficient.

More precise than the guide. The guide says a coordinator needs "Task" in allowedTools to invoke subagents. In the Agent SDK, allowedTools only pre-approves tools, and the delegation tool runs without approval whether listed or not. Claude Code 2.1.63 renamed the tool Agent, and Task still works as an alias (checked 2026-09-30). Exam answer: include "Task" in the coordinator's allowedTools.

Interactive · which subagents does this query need?

Top to bottom: the coordinator makes one Task call per subtask to document analysis subagents split by source type and web search subagents split by subtopic; their final messages return to the coordinator; synthesis runs; the coordinator checks for gaps and either re-delegates a targeted query to web search along a dashed path or passes to report generation. Buttons switch between a one-fact question, a two-item comparison and a broad topic, showing which subagents are invoked for each. coordinator reads query · picks subagents one Task call per subtask document analysis · by source trade surveys legal papers not invoked web search · by subtopic visual query LFP music Na-ion writing film games design final messages only · back to the coordinator synthesis coordinator: any gaps? gap → targeted query coverage sufficient report generation coordinator answers directly 12 subagents · clearly divided responsibilities 1 web search subagent · 3–10 tool calls 2 web search subagents · 10–15 tool calls each

Choose a query to see which subagents the coordinator invokes.

Simple fact-finding: one web search subagent. Document analysis, synthesis and report generation are not invoked.

Direct comparison: one search subagent per chemistry. No long documents to read, so document analysis is not invoked.

Broad research: search split by subtopic, analysis by source type, and a gap check that can re-delegate before the report.

Time reads downward. A greyed lane is a subagent type the coordinator did not invoke. The dashed path is iterative refinement: a gap sends a targeted query back to web search, then synthesis runs again.

SkillSelect which subagents to invoke, not just how many

The coordinator analyzes each query and dynamically selects which subagents to invoke, rather than always routing through the full pipeline. For the year a standard was ratified, it needs one web search subagent making 3–10 tool calls, and it answers without document analysis, synthesis or a report. To compare two battery chemistries, it runs two search subagents, one per chemistry, and skips document analysis because no long documents need reading. Anthropic's figure for a direct comparison is 2–4 subagents making 10–15 tool calls each. Only a broad topic warrants more than ten subagents.

SkillPartition the scope so subagents cannot overlap

Subagents cannot see each other's assignments, so the coordinator prevents duplication by partitioning the scope into distinct subtopics or source types, one per agent. Anthropic's early runs gave several subagents the brief "research the semiconductor shortage", and two of them duplicated each other on 2025 supply chains. The fix was a four-part brief: an objective, an output format, guidance on tools and sources, and clear task boundaries.

SkillWorked example: too narrow a decomposition

Asked about "the impact of AI on creative industries", the system returns a cited report on digital art, graphic design and photography. Every subagent succeeded, and nothing says music, writing or film are missing. The coordinator's log shows three subtasks, all visual. That overly narrow decomposition left most of a broad research topic unassigned, so no subagent could notice the gap. The first fix is to list the sectors before assigning them.

Before

Three subtasks, all visual

// coordinator log, pass 1
subtasks = [
  "AI in digital art creation",
  "AI in graphic design",
  "AI in photography",
]
After

The topic listed, then split

// coordinator log, pass 1
sectors  = ["visual art","music","writing",
            "film","games","design"]
subtasks = one per sector, each with
           objective, output format,
           sources, boundaries

The second fix is the refinement loop. After the first synthesis pass, the coordinator checks the draft against its sector list. The film section cites one source, from 2019, so it re-delegates a targeted query to one search subagent instead of rerunning every search. It then re-invokes synthesis, and stops when a stated coverage bar is met, not after a fixed number of passes.

// coordinator, after synthesis pass 1
gap:  "film" cites 1 source, dated 2019
re-delegate → web search subagent
  objective:  "AI in film production 2023–2026:
               VFX, dubbing, script tools"
  output:     "5–8 findings as {claim, url, date}"
  sources:    "trade press, studio filings"
  boundaries: "film and TV only; other sectors are done"
re-invoke synthesis with all findings
stop when: every sector has 3+ sources from 2023 on
The decision test Did a component fail, or did every component succeed? Something failed → debug that component. Everything succeeded, yet the report is incomplete → fix the coordinator's decomposition and its gap check.
Takeaway
  • The coordinator picks which subagent types run, and how many. A one-fact question needs one search subagent.
  • Partition by subtopic or source type, because subagents cannot see each other's briefs.
  • Refinement: check synthesis for gaps, re-delegate a targeted query, re-run synthesis until a stated coverage bar is met.

check yourself

0 / 3 answered

Analysts now send the research system many quick questions, such as which company acquired a named startup and when. Every query runs web search, document analysis, synthesis and report generation, taking about three minutes. Quick questions must be answered in under 20 seconds, and broad topics must still get the full cited report.Which change best meets the 20-second target while preserving full reports for broad topics?

Select one.

For the topic "remote work's effect on city economies", the coordinator sends three web search subagents the same brief: "Research remote work and cities." Two return near-identical findings on office vacancy, and nobody covers tax revenue or transit ridership. Every subagent succeeded. Every exchange between agents must appear in the coordinator's log.Which TWO changes to the coordinator best fix both the duplication and the missing coverage?

Select two.

To save a round trip, an engineer lets the document analysis subagent send its findings straight to the synthesis subagent. A week later a report cites a figure that analysis had flagged as unreliable, and the coordinator's log shows nothing about the handoff. The coordinator must decide which flagged findings synthesis may use.Which change best restores control of the handoff while keeping analysis and synthesis as separate subagents?

Select one.

sources

  1. Anthropic Engineering, "How we built our multi‑agent research system": the scaling figures (1 agent with 3–10 tool calls, 2–4 subagents with 10–15 calls each, more than 10 subagents), the semiconductor-shortage duplication case, the four-part brief, and the lead agent deciding "whether more research is needed". Checked 2026-09-30.
  2. Claude Code Docs, "Subagents in the SDK": the Agent tool's prompt string is the only content passed to a subagent, only its final message returns to the parent, and the tool appears as Agent in tool_use blocks but Task in the system:init list. Checked 2026-09-30.
  3. Claude Code Docs, "Configure permissions": allowed_tools auto-approves listed tools, any tool not listed "is still available to Claude", and Agent is among tools that "don't ask before running". Checked 2026-09-30.
  4. The creative-industries decomposition failure, "which subagents to invoke", iterative refinement with targeted queries, and routing for observability: Claude Certified Architect — Foundations Exam Guide v1.0, Domain 1.2 and Sample Question 7. The battery and ratification queries and the film re-delegation are this course's examples.