Working with one AI assistant is fairly contained. You ask, revise, reject, and continue. The history of the work usually sits in one place.
Working with several changes the experience. Research happens in one conversation. Production happens in another. A third system reviews the result. Each can be useful on its own, while the work between them starts to accumulate.
I have been experimenting with this in creative projects, separating research and critique from image production and using different systems for planning and refinement. The additional capacity is real. So is the need to carry decisions from one place to another.
A direction changes. One agent needs the updated brief. Another needs to know why the previous version was rejected. The shared document needs to catch up.
There can be more help available and more work to keep that help aligned.
More Output Is Not the Same as More Productivity
The most visible evidence of AI productivity is often the easiest to count: drafts generated, research completed, variations produced. Each additional agent appears to expand the amount of work the system can do.
But a completed subtask has value only in relation to the outcome it supports.
Imagine a marketing team preparing a campaign recommendation. Three agents can research competitors, propose positioning, and develop creative directions simultaneously. That may shorten the wait for a first pass. If each uses a different audience definition, someone must reconcile the recommendations before the team can act.
The outputs are finished. The recommendation is not.
This distinction matters because elapsed time, human effort, and computational cost are separate measures. A workflow can become faster while consuming more resources. It can also cost more and still be worthwhile because it produces a substantially better decision.
In its account of building a multi-agent research system, Anthropic reported a 90.2% improvement over a single-agent baseline on its internal research evaluation. It also reported that multi-agent systems used roughly 15 times as many tokens as ordinary chat interactions in its data. That token comparison was against chats, not the single-agent research baseline.
Those findings describe one implementation, not a universal multiplier. They make the economic distinction visible: improved performance and improved efficiency are different claims.
The relevant measure is the value of the result relative to the total effort required to achieve it. Counting production alone leaves part of that effort outside the calculation.
Coordination Is the Cost of Dividing Work
It is tempting to assume that software should be inexpensive to coordinate. Agents can respond quickly, work in parallel, and exchange structured information. They do not need meetings to get started.
Yet coordination arises from the relationships between tasks.
Thomas Malone and Kevin Crowston's 1994 study of coordination defined it in terms of managing dependencies among activities. The useful unit of analysis is the dependency: what one activity needs from another before it can proceed successfully.
In my creative experiments, Seso handles research, curation, and critique, while Zey handles image production. I retain creative direction and final selection. Separating those responsibilities gives the work a clearer structure. It also creates dependencies. A critique needs to become a production instruction. A new image needs to be assessed against the direction already chosen.
These are small personal experiments, with some coordination still carried by me. They illustrate the work involved in a handoff; they do not establish whether the arrangement is more efficient overall.
The same distinction appears in controlled research. A study spanning 180 agent configurations found that the benefits of additional agents varied substantially by task and architecture. Some parallelizable tasks improved, while the tested multi-agent approaches performed worse on sequential planning. These are benchmark results, not fixed rules for every workflow.
The implication is narrower and more useful than “more agents create more overhead.” Costs depend on what must be coordinated, how often it changes, and how tightly one contribution constrains the next.
Independent research streams can proceed with relatively little interaction. A closely connected chain of decisions may require frequent reconciliation. Adding capacity to that chain can create new coordination work precisely where the original task was already difficult.
Context Is the Invisible Cost of Collaboration
A handoff can look complete because the file arrived and the instructions are readable. The receiving agent has the output. What it may lack is the history that makes the output meaningful.
Creative iteration makes this easy to see. A final image prompt records the composition, palette, texture, and distortions to produce. It may preserve much less about why those choices survived several rounds of selection.
There is a difference between knowing that an image should be less realistic and knowing what kind of realism was rejected. A variation might have been too literal, too polished, or too close to an established visual form. Those objections can produce very different revisions.
The same problem exists in marketing strategy. A positioning statement might be the result of rejecting a broader audience, a cheaper offer, and a more familiar category. Transfer the statement without those decisions, and a new agent may reasonably recommend reopening choices the team has already resolved.
The missing context is not necessarily a model failure. It may simply never have been supplied.
What needs to travel includes the task, the decision history, and the standard used to judge the result. Preserving all three takes effort. Preserving too much can make the important material harder to find. A useful handoff therefore requires selection, not a transcript dump.
This creates a cost that is easy to miss: reconstructing the reasons behind work that already exists. The artifact survives, while part of the investment that made it useful has to be made again.
Context loss should also be distinguished from legitimate disagreement. Two agents can read the same brief and reach different conclusions. Better documentation reduces avoidable divergence; it does not eliminate interpretation.
The economic value of context lies in how much unnecessary reconstruction it prevents while leaving room for a better decision.
Shared Memory Needs Rules for Its Use
The obvious response is to give agents a common place to store the work.
That is the logic behind the shared working library I have been building across AI tools. Project briefs, prompts, iteration histories, and feedback can persist beyond the conversations that produced them. A new session can begin with a record of prior decisions.
This is useful infrastructure. Its value depends on how it enters the workflow.
A document can be available without being retrieved. It can be retrieved without being current. It can be current without making clear which instructions take priority.
One agent's response during the experiment made the limitation explicit: changes in the shared brief would be picked up when it reread the document, not automatically. The file had become persistent. The update had not become synchronized.
Research shows that structured information sharing can help. In a multi-agent travel-planning study, researchers found that a shared notebook improved the tested system's ability to preserve retrieved details. Adding an orchestrator provided further gains. The study examined a specific planning environment; it supports the usefulness of explicit information-sharing mechanisms rather than proving that any shared folder creates alignment.
A repository holds information. Coordination also requires a way to identify authoritative instructions, surface changed decisions, and retire obsolete versions.
In practical terms, the system needs to distinguish what was proposed from what was accepted. Otherwise, a comprehensive archive can preserve contradictory directions equally well.
Some context work can be paid for once and reused. A maintained brief can reduce repeated explanation across many sessions. But maintenance is part of that investment. Without it, the burden moves downstream to every agent or person trying to determine what the archive currently means.
Evaluation Becomes a Constraint on Production
Even with clear responsibilities and current context, multiple agents can produce several persuasive answers.
Consider three agents reviewing a marketing strategy. One recommends narrowing the audience. Another prioritizes distribution. A third argues that the offer needs to change. All three can support their recommendations with coherent reasoning.
The difficult work is now identifying which constraint matters most.
These recommendations may address different levels of the same problem. They may depend on incompatible assumptions. They may also be alternatives whose relative value can only be established through a test. More reasoning does not automatically settle which situation applies.
Agreement offers limited reassurance when the answers are not independent. An ICML 2025 study of correlated model errors found substantial overlap in mistakes across models, including models from different providers. Several agreeing answers can therefore repeat the same error.
At the same time, multiple outputs can improve reliability. A study comparing debate and majority voting found that voting accounted for most of the gains attributed to debate across seven tested language benchmarks. Discussion itself was not the main source of improvement. That finding concerns tasks with assessable answers, rather than open-ended business decisions.
The distinction matters. A workflow benefits from additional perspectives when it has a credible way to assess them. That might involve checking claims against primary evidence, running a calculation, testing a prediction, or comparing work against a defined standard.
Asking another agent to judge can help, but it introduces another output whose reliability matters. The evaluation process needs grounding of its own.
This is where cheap production can create expensive accumulation. If generating recommendations becomes easier than resolving them, unresolved analysis grows faster than decisions.
For a marketing leader, the cost appears as attention spent comparing plausible alternatives rather than advancing the work. The scarce resource may become the capacity to determine what deserves action. Producing another recommendation has little marginal value when it does not change that determination.
The Organization Behind the Intelligence
A small team can now access kinds of analysis and production that once required more people, time, or outside support. That expands what it can attempt. It also changes the work required to make those attempts coherent.
The organizational opportunity is not confined to automating existing tasks. It includes reconsidering where work should be divided, where decisions should remain connected, and which knowledge deserves to persist.
Those choices shape whether new capacity becomes usable capacity. They also shape whether learning compounds. A decision preserved with its evidence can improve the next project. A decision buried among abandoned drafts may have to be rediscovered.
Senior work in this environment includes designing those conditions: giving decisions a clear owner, making standards inspectable, and arranging feedback so that new evidence can change the work.
The advantage worth building is an organization that can expand its capacity without repeatedly losing its understanding of what it is trying to accomplish.