Most descriptions of multi-agent work begin with the cast: a researcher, a strategist, a critic, a writer. That is understandable, because roles are easy to name and parallel activity is easy to show. But a list of agents is not a managed workflow.
The useful unit of design is the workflow contract. Before work begins, we define what must happen, in what order, what each stage must produce, what evidence must survive into the next stage, and the conditions under which an output is accepted, rejected, or run again. The orchestrator agent then manages that contract. It dispatches specialist agents and personas, checks their artifacts against the criteria, records decisions, and redirects work when the criteria are not met.
The animation below is the execution trace of one contextual-targeting study. The question was where a prenatal supplement could appear next to fertility content in Italy without feeling intrusive, and which advertising approaches each placement allowed. Press play to watch the run, pause on any stage, click a bar, or use the step controls. The upper rail is the orchestrator. The lanes are specialist functions. The growing area at the bottom is the work produced; red marks work the orchestrator rejected rather than allowing it to contaminate the study.
The workflow comes before the workers
A workflow should be strict where errors compound. It means fixing the interfaces between stages.
For each step, we define five things:
- Input boundary. Which files and prior decisions may this step use? An analysis agent should not quietly read a later creative recommendation, and a re-run should not ingest the material that was rejected.
- Required artifact. The step must return a checkable object: a segment map, panel specification, interview guide, transcript set, coding file, challenge memo, or final report—not a claim that the task is complete.
- Acceptance criteria. The artifact must have observable properties. Six segments means six. Line-by-line coding means every exchange has a source. Three routes per segment cannot be three phrasings of the same idea.
- Invariants and exclusions. Some decisions must persist through the entire run. In this study, content about pregnancy loss was excluded early. It stayed excluded from respondent generation, fieldwork, placement planning, and the final report.
- Failure action. The workflow says what happens when a criterion fails: retry the same task with a corrected input, redirect it to a different specialist, discard a branch, or stop the run.
This is what makes parallelism safe. Four agents can explore independently because their outputs meet at a defined consolidation step. Six simulated specialists can widen an interview-question bank because a later synthesis stage has an explicit job: remove duplicates and closed questions, preserve coverage, and keep product references out of the first three interview phases.
The orchestrator is not a project-status narrator
An orchestrator does not simply announce that one task is finished and start the next. It owns the run state and has the authority to make consequential decisions.
At the first stage of this study, four independent context analyses were dispatched. Three returned. The missing decision-influence analysis covered the partner, doctor, pharmacist, family, and forums involved in the decision. The workflow did not permit the study to continue with that perspective absent, so the orchestrator re-issued it. That recovered analysis later produced the study’s central finding: the customer was surrounded by people with partial authority and partial information, leaving a gap a brand could fill in function but never credibly claim as a personal relationship.
The point is not that a retry occurred. Retries are trivial. The point is that the orchestrator knew why the missing artifact was structurally necessary and refused to silently reduce the scope.
Later, all 36 interviews were completed. They produced 220,651 words. They also failed acceptance: some respondents saw an English stimulus and others an Italian one. A reaction to language could no longer be separated from a reaction to placement. The output was large and unusable.
The orchestrator discarded the entire set and re-ran fieldwork with one variable corrected and everything else held constant. The rejected transcripts remained outside every downstream agent’s input boundary. This is orchestration in its least glamorous and most valuable form: protecting the integrity of later work by refusing completed work that does not meet the contract.
Agents perform methods; personas provide situated evidence
We use agents and personas for different purposes.
Agents perform bounded methods. A behaviour analyst reconstructs the actual journey. A sensitivity analyst grades where advertising becomes intrusive. A transcript coder works line by line. A category analyst preserves counter-evidence. A boundary tester pushes a creative route until it fails. Their roles are defined by operations, inputs, and outputs.
Personas occupy specified situations. The 36 synthetic respondents in this run were not a cloud of generic “Italian women.” They were allocated across six contexts, with contrast variables inside every segment and stable language and form-of-address rules. Personas were research instruments inside the workflow; they did not decide what the workflow should be.
Giving a persona control of method confuses simulated experience with analysis. Giving a method agent an underspecified persona encourages it to invent the population it is supposed to study. The orchestrator keeps both inside their proper boundaries.
Acceptance criteria are where strategy becomes executable
Vague instructions defer judgment until the end, when errors are hardest to locate. “Find good placements” sounds strategic but gives an orchestrator nothing to test. The executable version is closer to this:
Select six contexts supported by more than one independent analysis, with named inventory that can actually be bought, deliberate coverage across sensitivity levels, and an explicit exclusion list carried into every later stage.
Likewise, “develop creative ideas” became four parallel creative approaches across all six segments, each route requiring a headline, body copy, placement, supporting interview quote, and stated risk. A separate challenger then tested the assumptions shared by those creative agents. Only after the challenge was on file did a selection agent choose three genuinely different routes per segment and record whether each challenge was accepted, addressed, or rejected with a reason.
These criteria do more than check quality. They determine the architecture of the work. Independence belongs before convergence. Expansion belongs before compression. Challenge belongs before selection. Traceability belongs throughout.
A trace is part of the deliverable
The final report matters, but so does the ability to explain how it came to exist. This run passed through 16 controlled stages over 4 hours and 31 minutes, producing 983,208 words before rejected work was removed and findings were compressed into a 13,791-word client report.
The final claims can be traced backwards: from a recommendation to a category, from a category to coded exchanges, from those exchanges to a specified respondent and interview condition. Exclusions and limitations travel forward with the findings instead of disappearing during polishing.
The animation makes that management visible. You can see parallel dispatch, consolidation, redirect, rejection, re-run, adversarial review, and final integration. It shows why the orchestrator sits above the lanes: not because it is the smartest persona in the room, but because it is responsible for the contract connecting every room.
The operating rule
Define the workflow strictly. Give every stage a concrete artifact and acceptance test. Separate method agents from situated personas. Give the orchestrator enough authority to inspect, reject, redirect, and stop. Preserve a source trail for what survives and a decision trail for what does not.
Then let the agents be creative inside the boundaries that make their work usable.