How several specialised assistants solve one task

A supplier moves a delivery date and sends a revised agreement. The business needs to recover earlier commitments from correspondence, compare the two documents, check the relevant rules and prepare a clear response. This looks like one request, but it contains several different kinds of work.

Putting every tool and permission behind one assistant would make the interface simple while making responsibility obscure. We designed the system so the user keeps one conversation, but each internal step has a defined owner, a limited set of materials and a result that can be checked.

Giving one assistant search, documents, video, legal sources and access to business systems seems convenient. Yet every added capability expands the possible error. It becomes difficult to explain why a source was chosen, what data was visible and whether an action was permitted.

We separated the responsibilities. A general assistant leads the conversation, a legal role works with legal materials, and a research role handles tasks that require many sources. A coordinator chooses the right role for each step and assembles the result.

The coordinator separates correspondence review, document comparison, legal research and preparation of the final response. It also gives those parts a sensible order: there is no point assessing a changed clause before confirming that the latest document is actually present.

The user remains in one conversation instead of carrying results between separate tools. The internal discussion is not dumped into the interface, but the plan and source trail remain available for review.

A legal assistant needs legislation, judgments and document templates, but not a media archive. A researcher needs search and page reading, but not permission to send messages or alter business records. Narrow permissions limit the consequences of a bad decision.

Users are separated as well. Files, conversation history and working results do not mix. Even a useful tool should see no more information than the current task requires.

Research may cover dozens of pages. If an external source fails near the end, reading every completed page again wastes time and can change the result. Completed steps are stored so work can continue from the point of interruption.

Retries are limited. Temporary failures can be attempted again, but the system must not repeat the same mistake forever. When progress needs a new decision or permission, the task stops and returns a clear explanation to the user.

A specialist can be wrong, and the coordinator can combine correct fragments incorrectly. Important claims therefore keep links to the material behind them. If reliable sources disagree, the disagreement appears in the result rather than being smoothed away.

A legal draft is not advice merely because it sounds assured, and a research summary does not replace its primary material. The system organises the work; people retain the final decision in consequential situations.

When a search source, document format or business rule changes, one role can be updated and checked without rebuilding the whole service. Other roles continue operating under their existing rules.

Quality can be tested in the same way: legal work is checked for accurate sources, research for coverage, and coordination for correct assignment and ordering of steps.

From the outside this remains one assistant. The user states a goal, answers necessary questions and receives a combined result. Behind it is a trace of who performed each step, which information was used and why work stopped or changed roles.

The value does not come from the number of assistants. It comes from understandable responsibility, limited access, preserved intermediate work and the ability to trace an important conclusion back to its source.