A digital workforce should expand when independent work can justify the cost of another Worker, and contract when coordination becomes the bottleneck. The same mission may need one Worker to resolve an ambiguity, several to pursue independent tasks, and one accountable owner to bring the result together.
We call this direction Dynamic Workers: adapting the division of work, the size of the team, and its responsibilities as a mission develops. Our hypothesis is that deciding how to organize the work should become part of the system's intelligence.
The consequence is significant. A customer could delegate an outcome without having to design and manage the team that delivers it.
Another Worker needs somewhere useful to work
The number of available agents says little about how much faster a mission can finish. Useful capacity depends on what can advance independently, what must happen in sequence, and what the Workers share.
Four constraints matter:
- Independent work. Separate sources or records can be investigated concurrently when each contribution can be checked and combined. Additional Workers help only while useful independent work remains.
- Dependencies. A decision required by every subsequent step can hold the whole mission back. Adding Workers downstream does not resolve the missing decision.
- Shared resources. A session, document, or application may be a point of contention. More Workers can introduce conflicting edits, duplicate actions, or waiting.
- Coordination. Assigning work, transferring context, resolving disagreements, and verifying the combined result all consume effort. That effort belongs in the cost of delivery.
The practical question is where the next Worker can remove a constraint. If the constraint is a decision only the customer can make, unclear authority, or a required approval, increasing the team may accomplish nothing.
The bottleneck moves during the mission
Even a well-chosen team can become a poor fit as a mission develops.
An investigation may begin with several independent leads. Most may close quickly, leaving one difficult dependency. A resolved uncertainty can then reveal several new tasks that can proceed at once. The amount of useful parallel work changes even though the customer's goal stays the same.
An orchestrator should therefore be able to divide work, bring contributions together, and retire assignments that no longer advance the mission. It should also recognize when a human decision would unlock progress more effectively than another agent.
Reassessing and reorganizing the team has its own cost. Improvements in completion time, quality, or human effort must justify that cost. A stable team may remain the better choice. Adaptation must earn its place through improved delivery.
Delegation must carry responsibility
A delegated task needs more than an instruction. It needs a defined contribution, the context necessary to make that contribution, clear permissions, and evidence that it was completed.
Correct parts can still produce an incorrect whole. Two Workers may report accurate prices under different contract terms. Combining those prices without preserving the conditions creates a misleading recommendation. Verification must cover the relationships between contributions as well as the contributions themselves.
Shared state makes responsibility more consequential. A Worker editing a customer record must know the limits of its authority. Another Worker must be able to distinguish a completed change from an action still awaiting confirmation. Ambiguity here can turn parallel effort into repeated or conflicting work.
Someone must own the final result. The customer should receive a coherent deliverable, the evidence behind it, and unresolved decisions. If they must reconstruct the answer from agent outputs, the mission remains unfinished.
Measure the mission through completion
Dynamic Workers should be judged against both a capable solo Worker and a fixed team. Beating the solo baseline alone would not establish that adaptation helps; ordinary parallel execution might explain the improvement.
The comparison needs four standards:
- Equivalent completion. The same scope, coverage, and verification requirements apply to every approach. Missing evidence is unfinished work.
- Time to a usable result. Include waiting, retries, reconciliation, and final checks. For time-sensitive work, a correct result delivered too late can still fail the assignment.
- Full delivery cost. Include execution, coordination, infrastructure, recovery, and the attention supplied by a person. A faster team may be more expensive.
- Dependability. Examine failures and slow runs alongside typical performance. A system that requires frequent rescue transfers work back to the customer.
Different commitments require different judgments. A deadline demands timely completion. An ongoing responsibility demands continuity and an appropriate response when conditions change. Worker count alone measures neither.
What the research must establish
Our starting point is a preliminary conceptual framework for browser-agent scaling. It identifies how workload, dependencies, and coordination can limit useful concurrency. It reports no empirical scaling results and does not establish that adaptive teams outperform fixed ones.
Dynamic Workers extends that framework into a product and research hypothesis. Our ambition is to carry it across applications, screens, and devices, from solo execution to coordinated teams and swarms. Those broader capabilities remain work to establish.
The decisive result would be better verified completion at an acceptable total cost, with less effort required from the person delegating. Knowing when to use fewer Workers belongs in that result.
The customer delegates the outcome. Organizing the work should be the workforce's job.