Parallel vs sequential orchestration: choosing a topology
AI-assisted research. This note was prepared with the assistance of AI tools. It may contain errors and is not professional advice. See our Legal Disclosures.
As marketing programs connect more agents, the arrangement of those agents starts to matter as much as their individual skill. The same four agents can finish a job in minutes or hours, cheaply or expensively, depending on how the work flows between them.
That arrangement is often called the topology. It shapes how fast work completes, what it costs, and how easy it is to understand when something goes wrong. Two basic shapes, sequential and parallel, cover most marketing workflows, and most real systems combine them.
Sequential orchestration
In a sequential design, each step waits for the one before it. A researcher gathers facts, then a writer drafts, then an editor revises, then a checker verifies. Our post on multi-agent content pipelines describes this shape in detail.
The strengths of the sequential shape come from its order. Each step builds on the previous output, so later steps have more to work with. The flow is easy to follow and easy to debug, because there is one path to trace. Errors can be caught at each stage before they spread to the next.
The weaknesses come from the same place. Steps cannot overlap, so total time is the sum of every step. A failure anywhere blocks everything after it, and a slow step holds up the whole chain.
Sequential design fits workflows where later steps genuinely depend on earlier ones. Content pipelines, planning followed by execution, and research followed by analysis are the common cases.
Parallel orchestration
In a parallel design, independent tasks run at the same time, and their results are combined afterward. A planner might send 10 account research tasks to 10 agent instances at once, then compile the briefs into one report.
Parallel work is much faster for independent tasks, since total time is set by the slowest branch rather than the sum of all of them. A failure in one branch does not block the others, so you can retry a single branch without rerunning the rest.
The costs show up at the edges. Combining results can be hard when branches produce inconsistent formats or conflicting conclusions. Spending can spike, since many tasks run at once. Rate limits on tools and APIs are easier to hit, which our post on scaling agents for traffic spikes covers.
Parallel design fits independent tasks of the same type. Researching many accounts, generating variants for different segments, and checking many pages are good examples.
The merge step is where parallel work fails
Fanning out is easy. Bringing results back together is where most of the design effort belongs. Branches that each did their job well can still disagree with each other, because each worked without seeing the others.
We suggest designing the merge step before you design the branches. A few rules help.
- Fix the output schema: every branch returns the same structure, so the merge step can compare like with like.
- Share the source facts: every branch reads from the same brief and the same approved facts.
- Check for conflicts: the merge step looks for contradictions across branches, not just errors within each one.
- Report missing branches: if one branch fails, the merged result says so instead of quietly presenting a partial set.
The last rule matters more than it looks. A merged report that drops one failed branch but presents itself as complete is a silent truncation, and it is easy to miss.
Hybrid shapes
Most real workflows combine both shapes. A common pattern for a campaign launch runs in four stages.
- Sequential: plan the campaign and review the plan.
- Parallel: produce assets for each channel at the same time.
- Sequential: review all assets together for consistency.
- Parallel: schedule and launch across channels.
The review in the middle matters. When assets are produced in parallel, they can drift apart in message or tone. A consistency review catches this before launch. This is the same pattern our orchestration run diagram shows, with parallel execution followed by a single check step.
A worked example
Consider a hypothetical account-based campaign with 40 target accounts. Each account needs a research brief, and research for one account does not depend on any other. Suppose each brief takes a research agent a few minutes.
Run sequentially, the 40 briefs take 40 times as long as one. Run in parallel, they take about as long as the slowest brief, as long as the tools and rate limits allow it. Here parallel is the clear choice for the research stage.
The next stage is different. A strategist agent ranks the accounts and picks the 10 to pursue first. Ranking needs every brief, so it must wait for all 40 to finish. That stage is sequential by nature.
The outreach stage can fan out again, with one drafting task per selected account. A final consistency review reads all 10 drafts together before anything is sent. Our post on agentic ABM covers this workflow end to end.
How to choose a shape
The choice comes down to dependencies, deadlines, and the cost of combining results. We suggest working through these checks for each workflow.
- Dependency check: confirm whether each step needs the previous step's output, or only the original brief.
- Deadline check: decide how quickly the work must finish and whether a sequential chain can meet it.
- Merge check: name how you will combine and check results from parallel branches.
- Capacity check: estimate the rate limits and spending that apply if many tasks run at once.
A step that needs only the original brief is a candidate for parallel work. A step that needs another step's output is not, however tempting the speed gain.
What to measure
Topology choices show up in a few measures. Track them per workflow, so you can see whether a change in shape helped.
- End-to-end time: from trigger to completed output.
- Cost per run: model and tool spending for the whole workflow, as in our approach to cost per task.
- Merge conflicts: how often branches disagree in ways the merge step must resolve.
- Blocked runs: how often a single failure stops a sequential chain.
Open questions
We are studying when a workflow should change shape at run time. A planner could choose parallel research for a large account list and sequential research for a small one. That flexibility saves time, but it makes runs harder to compare, and we want to know whether the gain is worth the loss of predictability.
How to apply this: sketch your most common multi-agent workflow as boxes and arrows, and mark which steps actually depend on each other. Treat the independent steps as candidates for parallel work, and design the merge and consistency review before you split anything.
Related research
Supervisor agents and the limits of agents managing agents: judgment stays with people
Supervisor agents can coordinate and check other agents. They work best deciding what people should look at, and worst when they replace the looking.
Event-driven orchestration: triggers, queues, and retries for marketing workflows
Marketing work starts with events: a form fill, a launch, a drop in performance. Agent systems built around events respond faster, and queues keep them safe.
Orchestrating across channels: one plan, many executors
Customers meet a brand across email, ads, social, sales, and the website. Orchestration keeps each channel's agents telling one story from one shared plan.
Put this research to work
Marketix turns these patterns into software: agents, automated workflows, and a hub that keeps people in control. Sign up for early access.
