An agent can use tools and carry out several steps toward an outcome. Start with one assistant and a practical workplace deliverable: a campaign brief, source-based content drafts, or a weekly performance summary. Business intelligence means using data to understand performance and inform decisions. AI can help prepare that work; you still define the question and check the evidence.
What you will leave with
A one-page campaign or performance meeting brief with traceable claims, defined metrics, visible unknowns, and a reusable handoff.
Step 1
Name the marketing or reporting decision
Choose a question such as which campaign deserves further investigation, what message a launch should communicate, or what changed in this week's results. Name the audience and the decision the deliverable will support.
Supply approved inputs: a campaign brief, a published article, or a sample performance export. Identify the source and reporting period. Use only information your organization permits this assistant to process, and remove unnecessary personal fields. Start with sample data when learning the workflow.
Try this instruction
Prepare a one-page [campaign or performance] meeting brief using [approved sources]. Answer [decision question]. Include sourced facts, metric definitions, open questions, and recommended discussion points. Mark missing information. Do not invent results, publish content, change campaign settings, or modify the originals.
You can name the decision, approved inputs, reporting period, and expected deliverable.
Step 2
Get one complete result from one agent
For marketing, ask for the audience, message, supporting evidence, channel, and next action. When repurposing an article into email or social drafts, require every factual claim to remain grounded in the supplied source. Do not let the assistant invent customer quotes or performance promises.
For business intelligence, check missing values, duplicates, date ranges, currency, and filters before summarizing. Define each KPI and its calculation. Recalculate at least one important total or rate independently; a polished chart is not proof that its numbers are correct.
Claims trace to sources, metrics have definitions, and a spot-check reproduces an important result.
Step 3
Delegate only work that can run separately
A lead owns the brief. A researcher gathers approved evidence. A writer drafts campaign material. An analyst checks data and calculations. A reviewer challenges claims and conclusions. These are roles; one assistant can perform them in sequence.
Independent source research and data-quality inspection can run in parallel. Reviewing a finished report depends on that report existing. Give the reviewer the sources and formulas, not just the analyst's summary. Extra agents consume more tokens and add coordination work, so a routine task may be simpler with one.
Each delegated task has an independent purpose and a result the lead can evaluate.
Step 4
Give each helper a clear handoff
Include the outcome, sources, allowed actions, owner, dependencies, and acceptance checks. For a weekly report, record metric definitions, filters, reporting dates, and the location of the approved export. Keep that context in a project document.
Give the final output one writer and let one agent operate a shared desktop window at a time. For independent coding changes, worktrees separate project files; they do not create separate databases or permissions. Shared spreadsheets and external records also need clear ownership.
Try this instruction
Analyze [approved export] for [question]. Define the metrics and reporting period first. Return the calculation method, three supported findings, data-quality issues, and uncertainty in [output]. Do not modify the source or final brief. Give the reviewer enough detail to reproduce the results.
Every output has one owner, and helpers know where to put their findings.
Step 5
Check what your app actually connects
Codex supports native subagents that report to the main chat. Claude Code supports subagents; its separate agent-teams feature is currently experimental and disabled by default. Follow the documentation for the product you use.
Separate AI apps do not become a coordinated team simply because they are installed together. Provide an explicit handoff or use a verified integration. Shared files are not shared conversation history. Desktop also does not mean local-only execution: current Cowork sessions run in the cloud and reach permitted local resources through the desktop app.
You know how context is passed, where outputs are saved, and which resources the task can access.
Step 6
Close the loop and reuse the process
Ask for the output location, source dates, checks, uncertainties, and next action. A weekly performance report should distinguish an observed change from a proposed explanation; correlation alone does not establish a cause. Review campaign drafts before publication and recommendations before changing spend.
Save the brief, KPI definitions, and review checklist. Record actual decisions after the meeting. Prefer structured integrations for repeatable data work, and add recurring automation only after the inputs, outputs, and allowed actions are predictable.
You can find the result, explain its limits, and repeat the workflow without reconstructing the whole conversation.
When you get stuck
Two agents overwrite a document
Pause the edits, preserve both versions, choose one writer, and reconcile the differences before resuming.
A confident claim has no source
Ask for supporting evidence and label the claim unverified until you can check it.
The task reaches the wrong app or files
Stop it, restate the permitted scope, and review access. Do not solve confusion by granting blanket permissions.
Terms explained
- Agent
- An AI system that can use tools and work through steps toward a goal.
- Handoff
- A compact brief transferring a task, context, boundaries, and expected output.
- Subagent
- A helper started by a supported agent system for a bounded task.
- Worktree
- A separate checkout of a Git repository for parallel work.
- KPI
- A key performance indicator with a defined calculation and decision-making purpose.
Sources and further reading
The workflow and templates are teaching guidance. The sources below support the underlying tools and concepts; product setup and availability can change.
Take it into your next task
Your completion checklist
0 / 6 checked
Checks are saved on this browser. Mark them only after you inspect the evidence. These are your checks, not an automated certification.