Parallel work needs independent ownership

Parallel agents help when useful work can proceed independently. They cause trouble when each worker changes the same artifact or assumes that another worker checked a claim. The coordinator's first job is to define ownership and the merge contract. "Research everything" is not a bounded assignment.

Give each worker a question, permitted sources, output shape, stopping rule, and destination. Assign disjoint files or read-only outputs when possible. Require each factual claim to retain its source and uncertainty. The coordinator can then compare evidence rather than choose whichever prose sounds more confident. A worker's completion message is not a substitute for inspecting its artifact.

Some tasks have a dependency graph. If a pricing calculation needs a product inventory, the inventory should finish first. Starting both workers does not remove the dependency; it creates guessing. Parallelize independent branches and keep dependent transitions sequential. Distinguish duplicate review, where two workers deliberately inspect the same evidence, from accidental duplicated work.

Merge conflicts are sometimes semantic rather than textual. Two workers may produce different but individually valid definitions of "active user". Their tables will merge cleanly yet produce a false comparison. Specify shared definitions before delegation. If evidence conflicts, preserve both claims and examine their dates, populations, and measurement methods. Majority voting cannot repair a systematic shared error.

A fictional agent reviews three vendor integrations. Worker A inspects authentication, B studies rate limits, and C studies data export. Every result includes vendor ID, source URL, checked date, supported claim, and open issue. The coordinator owns the final comparison.

A says vendor V2 supports organization tokens. C says exports require user tokens. These claims need not conflict because the endpoints differ. The coordinator records the scope explicitly and refuses to compress both into "V2 supports tokens". B cannot access one rate-limit page, so that cell stays unknown. The final artifact has two verified comparisons and one documented gap, not a fabricated complete score.

You have two workers and a task to compare five database products. One worker proposes writing the full final report while the other writes a competing full report. Design a better split for limited time.

A model solution assigns products 1 through 3 to one worker and 4 through 5 to the other, with the same criteria and evidence format. The coordinator normalizes definitions and checks a sample from each. If a particular criterion requires specialist review, a criterion-based split is also valid, provided the merge schema is fixed. Award one point each for bounded ownership, shared definitions, evidence retention, and a merge review. Do not award credit merely for using more workers. Parallelism is useful only when it reduces elapsed work without hiding responsibility.

Inspect this fictional assignment. It gives a worker enough context to act independently while leaving the merge decision with the coordinator.

1worker: vendor-auth2question: Compare authentication for vendors V1, V2, V3.3scope: Read public source-owner documentation only.4output_rows:5  vendor_id, endpoint_scope, auth_method, source_url,6  source_date_if_stated, checked_date, claim, uncertainty7ownership: auth-evidence.md only8stop: six opened useful pages or 20 minutes9must_not: infer price, send messages, edit final report10completion: rows plus explicit gaps and rejected claims

The output schema includes endpoint scope because "supports OAuth" may be true for one API and false for another. It includes uncertainty because inaccessible or ambiguous documentation must not turn into a confident yes/no. File ownership avoids two workers editing the same final artifact.

The coordinator still has work to do. It checks that the output exists, that sources support the claims, and that missing rows are accounted for. Delegation transfers a bounded task, not responsibility for the final conclusion. A worker can finish its assigned search while the overall comparison remains incomplete.

Suppose worker A defines the product taxonomy while worker B classifies products under that taxonomy. B cannot complete a valid classification before A's definitions are available. Starting both in parallel encourages B to invent provisional categories. The better design either finishes the shared taxonomy first or gives B a genuinely independent task such as collecting raw product descriptions.

This dependency table makes useful concurrency visible. The first two tasks can run together. Classification and recommendation remain sequential. More simultaneous workers would not remove their information requirements.

Two workers may cite different versions of a provider page. One says requests are limited per minute; another says the limit is per second. Before averaging or choosing, inspect the endpoint, account tier, publication date, and exact wording. The statements may describe different scopes or one may be obsolete. Preserve the rejected interpretation and the reason in the working notes so the same confusion does not return later.

A second independent reviewer can help find mistakes, but agreement is weak when both workers use the same incorrect source. Source diversity and explicit checks matter more than the number of agreeing agents. The merge owner should test a few critical claims directly, especially those that drive a consequential recommendation.

"No text conflict means the outputs agree" ignores semantic differences in units, definitions, and populations. A clean concatenation can still create a false comparison.

"Three agents agree, so the fact is verified" ignores shared sources and shared assumptions. Agreement without independent evidence can reproduce one error three times.

Two workers estimate migration effort. A reports three days excluding data cleanup; B reports five days including cleanup. The coordinator wants to average them to four days. Repair the comparison.

The solution first aligns the scope. It either separates migration and cleanup into distinct rows or asks for both estimates under one common definition. It preserves uncertainty and assumptions about data volume. The raw average is not meaningful because the estimates cover different work. Score one point for scope alignment, one for retaining the original assumptions, one for avoiding false numerical precision, and one for a next question that would make the estimates comparable.

Original practice: Two agents disagree about an API limit. How do you resolve it? A strong answer checks endpoint scope, plan, dates, and primary evidence. Follow up with both agents citing the same obsolete page. A weak answer trusts the majority or the longer response.