We will create a company-research assistant for a frontline team. Start a project and record its boundaries: research drafts with sources, while people retain responsibility for account-opening decisions.

Upload anonymized notes. Ask the assistant to separate entity confirmation, source checks, and comparable-sample rules. Keep scripts separate with documented inputs and outputs so the method can be maintained.

Create the research Service with a connection, saved prompt version, and allowed documents. Enable only required scripts and web access. Save a draft before testing it as a user.

In the project conversation, open Test a Service and select the draft. Replies now use its configuration. Administrator contact is simulated, but other authorized tools may still perform actions.

Open Tracing to inspect recorded file access and tool actions. It helps check evidence for verification steps, but a missing record does not prove that no action occurred.

Exit testing and return to the DeepAgents admin conversation. Review the missed verification and update the shared rule file: even a name-only request must start with background checks.

Restart testing with a fresh question. Verify the entity, year, and standalone reporting scope before research. If Service configuration changes, exit and reopen test mode.

After reviewing the test, return to My services, verify the prompt and access scope, and publish. Share the Service link so colleagues can begin their research tasks.

A colleague submits the synthetic case. The assistant confirms the entity, then shows sources, comparable-sample scope, and unresolved facts. Its output is a research draft for human review.

The colleague points out that Sample A uses consolidated group accounts. The assistant explains the mismatch, excludes that sample, and revises this report while retaining unresolved items.

The colleague suggests making entity and reporting-scope checks mandatory. With their confirmation, the Service submits feedback and context. This illustrates the published flow, distinct from simulated contact during testing.

Enable Service records so the admin can review feedback. A person reviews the change, then saves shared rules through DeepAgents or the file editor and retests. Feedback does not rewrite rules automatically.