You have two agents work on a report and get two complete reports back. What is missing now is not text but a decision between contradictory versions. There was parallelism; a sensible division of labour, not yet. The benefit begins where the subtasks deliver different results that can be merged.
Claude Code subagents work with their own context, their own assignment and configurable tools. Their use counts towards the same limits. An additional context can make responsibility clearer, but it does not provide a second, independent truth. Both agents can misunderstand the same source.
Splitting the assignment by what comes back
Our own example is a short report on a repair event. The sources are a schedule note, a materials list and confirmed decisions. Agent A is to write a clear draft. Agent B checks three given statements: room approved, number of participants confirmed and material costs approved.
Agent B is explicitly not asked for a second report. For each statement, its return consists of status, source location and justification. This forces the split into a usable interface. The main task can correct a specific condition instead of merging two texts by gut feeling.
{
"aussage": "Material costs approved",
"status": "open",
"quelle": "entscheidungen.txt, Material section",
"begruendung": "Approval is still a precondition."
}This self-designed format is one possible results contract, not a claimed native agent API. A text list with the same fields would be just as suitable for getting started. What matters is that “open” is not silently reinterpreted as “false” or “confirmed”.
The evidence reader does not need a writing role
The second agent is to read the named sources and return source locations. For that, it needs neither parallel editing of the report document nor the ability to send messages. Choose its tools to suit the task and also take the permissions of the environment into account. A tool list and actual access control are different layers.
Start with artificial files in a manageable project. That makes a missing source easy to spot. If Agent B asks for a file that was not provided, the return should name the gap. Missing access does not mean that the decision in question never existed.
Which work can really be parallelised?
Agent A can write from a clearly labelled state of the sources while Agent B checks the three statements. As soon as that state changes, both need a recognisable reference to it. A materials approval confirmed after the fact must not end up in the final version as both an open question and a settled fact.
So define one person or one main agent as the owner of the final file. The others provide notes or separate results. If everyone may change the same file, a content conflict becomes a write conflict as well. More parallelism does not resolve this dependency.
An interim update should also stay specific: files read, statements already checked and sources still missing. A blanket “almost done” does little to help the merge if the decisive source location is exactly what is missing.
Merging is a step of its own
Take the draft report and the evidence list together. For every discrepancy, go back to the original passage. If it says “after budget approval”, this condition must be preserved in the wording. Two agents agreeing with each other do not outweigh the source.
A useful final check also asks whether the report is understandable without the entire agent conversation. Are the sources recognisable? Have open responsibilities been preserved? Is it clear which person approved the final version? These questions concern the joint result, not how polished a single return looks.
The division described here was not run as a session. It shows a reproducible working contract: different results, limited tools, a recognisable state of the sources and exactly one place for merging. This contract alone already helps you use additional agents for a specific purpose.
Sources and documentation status: 25 October 2026.
