Multiple agents help when a job contains independent pieces. They also create more decisions, duplicated work and opportunities to overwrite something. Start with a lead and one helper. This walkthrough uses a weekly marketing performance brief: one agent checks a sample campaign export while another researches public market context. The method also works for code, research and documents. It is a proposed workflow you can practice, not a claim that every AI app shares the same team features.
What you will leave with
A small agent team produces one reviewed artifact, with a clear owner, sources, decisions and a handoff another person can follow.
Step 1
Decide whether a second agent helps
Write the final output first: a one-page performance brief explaining changes in spend and qualified leads. Checking a sample export and researching public market context can happen independently. Interpreting the results depends on both, so it happens later. For a short rewrite or one small bug, a single agent may be quicker. Add a helper for a bounded piece of work, rather than assigning impressive-sounding roles without a deliverable.
Each proposed helper has an output that can be completed without waiting for another unfinished task.
Step 2
Give the lead a compact shared brief
Put the outcome, audience, source files, constraints and definition of done in a short file. Include where outputs should go and which decisions remain yours. A folder can hold shared files, but it does not automatically give separate chats the same history. Ask the lead to pass the relevant brief into each delegation. Product behavior varies: Claude Code subagents have their own context, while its experimental agent teams provide a different coordination model.
Try this instruction
Own the final marketing performance brief. Use the provided sample export, metric definitions and public sources. Delegate independent research if supported. Give each helper its inputs, output filename and checks. Return sources and unknowns. Reconcile totals and date ranges. Separate evidence from hypotheses. Do not change campaigns, budgets or send messages. If native delegation is unavailable, give me a handoff I can paste into a separate chat.
A new helper can understand its assignment from the brief and named inputs alone.
Step 3
Assign one writer to each output
Let the analyst write metrics-check.md and the researcher write source-notes.md. Only the lead edits final-brief.md. The reviewer leaves findings in review.md rather than rewriting the final file simultaneously. Keep the ownership map visible. If two agents need the same file, make their changes sequential. For code, separate Git worktrees give parallel work separate working directories; they still need deliberate integration and do not isolate every external service.
No two active assignments claim the same editable file or production resource.
Step 4
Agree on a useful handoff format
A helper should return findings, supporting links, the file it wrote, unresolved questions and checks performed. Ask for source dates where freshness matters. Avoid giant transcripts that force the lead to repeat the investigation. For numbers, include the source, date range, currency, metric definition and denominator. Reconcile totals before comparing platforms. Separate confirmed facts from estimates. If a source is unavailable, the helper should say what could not be checked instead of filling the gap with a plausible answer.
Each important recommendation can be traced to an input or source, and uncertainties remain visible.
Step 5
Keep shared desktop control sequential
Two agents should not click or type in the same window at once. Assign a browser or desktop session to one operator while others research through independent tools or files. Tell the lead when a handoff changes the active window or application state. Where available, prefer a focused connector for a specific read over navigating an unrelated screen. Desktop permission is not blanket permission to email, purchase, delete or share private material.
You know which agent controls the visible app and which external actions are authorized.
Step 6
Integrate first, then review the whole result
After the inputs arrive, the lead resolves contradictory facts and writes the brief. The reviewer compares that actual artifact against the original requirements. Ask for specific defects: a total that does not reconcile, a changed metric definition or a claim that mistakes correlation for causation. Fix material findings and reopen the final document. For code, run checks on the combined change; separately passing branches do not prove that their integration works.
The final file opens, meets the brief and has no unresolved finding that blocks its intended use.
Step 7
Close the loop and keep the recipe
Record the final location, important choices, failed checks and next decision. Stop helpers whose work is complete and clean up temporary resources through the tool's supported workflow. Keep the brief and handoff format for the next reporting cycle. If coordination took longer than the work, simplify the team next time. Judge the system by a useful, correct result and the effort to get there, not the number of agents involved.
Someone else can find the result and understand what was decided without reading every chat.
When you get stuck
Helpers repeat the same research
Narrow each assignment to a different question and name the output. Make the lead responsible for overlap.
A helper misses an important instruction
Put that instruction in its actual delegation. Do not assume a separate context inherited every earlier message.
Two agents overwrite a document
Pause writers, preserve both versions, assign one integrator and reconcile against the latest saved content.
The team looks busy but never finishes
Reduce the task to one deliverable and a clear completion check. Add an explicit stopping condition for research.
Terms explained
- Handoff
- A compact transfer of the task, evidence, output location and unresolved work.
- Worktree
- A separate Git working directory associated with the same repository.
- Dependency
- Work that cannot be completed until another output or decision exists.
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 / 7 checked
Checks are saved on this browser. Mark them only after you inspect the evidence. These are your checks, not an automated certification.