Three agents agree on an answer. Have you gained three independent pieces of evidence? Only if their work adds independent support. Start by mapping the data and deliverable each worker contributes.
Before you begin: You understand a tool-using agent and shared application state.
An agent here is a model-driven component with a role, context, and permitted actions. Several agents can use the same underlying model. Giving them different names does not make their errors independent.
Find a task that can actually be divided
Suppose a research request needs information from three independent document collections. Separate workers can search those collections in parallel and return evidence. A coordinator can then combine results and identify gaps.
By contrast, asking a planner, thinker, critic, and writer to repeatedly discuss a simple factual question may add little. A single model with a read-only retrieval tool and an explicit output contract is a useful baseline.
Choose agent boundaries around different data access, tools, expertise requirements, or independently checkable deliverables. Boundaries based only on personalities are difficult to evaluate.
Define a handoff as data
A worker should return the task identifier, supported findings, source references, unresolved questions, and status. A coordinator should not have to infer completion from a paragraph that sounds confident.
| Field | Why it matters |
|---|---|
| Task and scope | Prevent work on a different question. |
| Evidence references | Let the next stage inspect support. |
| Status | Distinguish complete, partial, and failed work. |
| Open questions | Preserve uncertainty through the handoff. |
Pass only the context needed for the task. Copying the entire conversation and all credentials to every worker increases cost and access risk. Each tool still enforces the acting user's permissions.
What useful information does another agent add?
Several agents can repeat the same assumption and appear to agree. The value of an additional worker comes from a distinct contribution: another source, an independent calculation or a different responsibility with a clear handoff. More voices are not automatically more evidence.
Define the output of each role and how conflicts will be resolved. Preserve source links through aggregation so a final writer cannot turn disagreement into an unsupported consensus. The challenge asks whether the extra coordination produces better task outcomes after its time and cost are counted.
Three workers read the same incomplete source and repeat the same claim.
Give one worker an independent evidence check and require it to report contradictions.
The additional role can expose a shared assumption instead of merely increasing agreement.
Give shared state one clear owner
Parallel workers can overwrite each other's updates if they all edit a single mutable plan. Prefer separate results keyed by task ID and a coordinator that resolves conflicts deliberately. Use durable state when work must survive restarts.
A worker timeout does not mean its external action never happened. If a worker can change state, use operation identities and recovery procedures. Avoid giving both coordinator and worker responsibility for the same side effect without a clear protocol.
Two agents agreeing is not equivalent to two independent sources agreeing. They may share the same model, prompt, retrieved document, or misconception. A critic adds value when it checks evidence or a concrete invariant, not merely when it produces another opinion.
Remove one worker and predict what disappears
Suppose a coordinator has a catalog reader, a date checker, and a writer. The reader returns records; the checker validates date consistency against those records; the writer produces a response. Remove the checker and ask which failure should become more likely. If you cannot name one, its role may be too vague to evaluate.
How can a critic earn its place?
Give it a specific, testable responsibility such as detecting a date mismatch, and compare systems with and without that check on relevant cases. If ordinary code can perform the same check more reliably, use that as another baseline. A second opinion without a distinct evidence or validation boundary may only add cost.
Keep worker outputs separately so the coordinator can identify the source of a conflict. Then measure both total model usage and elapsed time. Parallel work can improve one while worsening the other; agreement alone measures neither.
Measure the system against a simpler baseline
Compare task success, unsupported claims, latency, total cost, and handoff failures. Parallel work can reduce elapsed time while increasing total model usage. Include unsuccessful tasks in cost accounting rather than reporting only cheap successful examples.
Exercise: three agents independently repeat a false fact found on the same webpage. Does majority voting increase confidence?
Compare your reasoning
Their agreement reflects a shared source, so it adds little independent evidence. Check the source and seek a genuinely independent authoritative reference. Record source provenance so the coordinator can detect this overlap.
Practice with feedback
Remove one worker and predict what disappears
A five-agent crew (researcher, analyst, writer, critic, manager) produces a newsletter. A single well-prompted agent produces a similar one.
Divide work only where the boundary helps, and compare against a simpler baseline.
Check your understanding
Your task
Ablate each role and decide what the crew should actually contain.
These notes stay on this page. Download them before leaving.
What to include
- The single-agent baseline is measured, not assumed
- Each ablation removes exactly one role
- Latency and cost are reported alongside quality
- Surviving roles are justified by parallelism, permissions, or tools
Compare with a worked answer
Here is one way to answer. Check how it uses the information in the task.
Task and fixed inputs: produce a 250-word member newsletter from a week of committee notes plus the current timetable. 20 fixed input sets. Baseline: one agent with a good prompt and both tools. Score 15/20 acceptable, 11s, 0.02 GBP.
| Configuration | score | latency | cost | what changed | |---|---|---|---|---| | full crew (5) | 17/20 | 48s | 0.11 | - | | minus critic | 13/20 | 36s | 0.08 | 4 factual slips returned | | minus analyst | 17/20 | 33s | 0.07 | nothing measurable | | minus manager | 17/20 | 29s | 0.06 | nothing measurable | | single agent | 15/20 | 11s | 0.02 | - |
Roles that survive: researcher and writer, because they need different tools (retrieval versus none) and the handoff between them is a structured claim list I can validate. The critic survives because removing it cost 4 points, and its job is checkable: it verifies each claim against its source id, which is a different operation from writing. The analyst and the manager produced articulate output and changed nothing. The manager in particular was routing messages between agents that could have been a function call. Removing both saved 19 seconds and 45% of the cost.
Final design: researcher -> writer -> critic, three roles, 29s, 0.06 GBP, 17/20. Against the single agent that is +2 points for 3x the latency and cost, which is worth it here because a factual slip in a member newsletter is expensive to correct after sending. Shared state and its owner: the claim list is owned by the researcher. The writer reads it and never writes to it; the critic appends verdicts to a separate verdicts list it alone owns. No structure has two writers.
When you are signed in, opening the challenge carries your edited working notes into its draft in this browser. The challenge has its own completion record. Practising here does not award points or mark it complete.
Next, build a research workflow whose outputs can be checked claim by claim.
Sources
The LangChain multi-agent documentation describes orchestration patterns. The A2A specification addresses communication between agents; neither source implies that extra agents always improve a task.