How agentic orchestration works

Our methodology in six diagrams: how we research, how the Agentic Hub coordinates agents and people, how work moves through a run, and how we decide what agents may do on their own.

The research loop

How every pattern we publish or build into Marketix moves from a narrow question to evidence.

01QuestionA narrow question thatevidence can answer02MapThe real workflow, data, andpeople03PrototypeThe smallest system that cananswer it04EvaluateTest sets, rubrics, and shadowruns05BuildTurn it into a productcapability06PublishFrameworks and findings,shared openlyEvidence logEvery step is recordedEach loop ends in a decision:continue, build, or stop
  1. 01QuestionA narrow question that evidence can answer
  2. 02MapThe real workflow, data, and people
  3. 03PrototypeThe smallest system that can answer it
  4. 04EvaluateTest sets, rubrics, and shadow runs
  5. 05BuildTurn it into a product capability
  6. 06PublishFrameworks and findings, shared openly
  7. LogEvidence logEvery step is recorded. Each loop ends in a decision: continue, build, or stop.
  8. Then back to step 01 with the next question.
The research loop. Each pass ends at the evidence log with a decision to continue, build, or stop.

Agentic Hub architecture

The coordination layer between people, specialist agents, and the systems they act on.

PeopleGoals and plansSet outcomes, priorities, and budgetsGuardrails and policyDefine limits, claims, and consent rulesApprovals and reviewApprove consequential work, handleescalationsAgentic HubCoordinates every agent and person against one planPlannerTurns goals into tasksRouterSends work to theright specialistShared contextPlans, facts,audiencesApprovalsTracks states,enforces reviewRun logRecords every decisionSpecialist agentsResearchAccounts, markets,sourcesContentDrafts, edits, checksDemandTargeting and outreachLifecycleJourneys and timingAnalyticsReporting and anomaliesTool connectors: scoped permissions, hard guardrails, every call loggedSystemsCRMCMS and siteEmailAdsAnalyticsKnowledge baseSignal data layerIdentity, consent, intent, and the unified event streamOutcomes and learning feed back into plans
  1. People
    • Goals and plansOutcomes, priorities, budgets
    • Guardrails and policyLimits, claims, consent rules
    • Approvals and reviewConsequential work, escalations
  2. Agentic HubCoordinates every agent and person against one plan
    • PlannerGoals into tasks
    • RouterRight specialist
    • Shared contextPlans, facts, audiences
    • ApprovalsStates and review
    • Run logEvery decision
  3. Specialist agents
    • Research
    • Content
    • Demand
    • Lifecycle
    • Analytics
  4. ConnectorsTool connectorsScoped permissions, hard guardrails, every call logged
  5. Systems
    • CRM
    • CMS and site
    • Email
    • Ads
    • Analytics
    • Knowledge base
  6. DataSignal data layerIdentity, consent, intent, and the unified event stream
  7. Outcomes and learning feed back into the hub's plans.
Agentic Hub architecture. People set direction and limits; the hub coordinates; agents act through governed connectors; outcomes feed back.

Anatomy of an orchestration run

How one piece of work moves through the hub, from trigger to learning.

PeopleAgentic HubSpecialistagentsTools andchannels01TriggerEvent orrequest02PlanTasks, owners,risks03Review planAdjust andapprove04DelegateRoute withcontext05ExecuteParallel wherepossible06CheckGuardrails,consistency07ApproveConsequentialwork08ActPublish, send,update09LearnMeasure andlog resultsEscalate when unsureResults inform the next plan
  1. Tools and channels01TriggerAn event or request starts the run
  2. Agentic Hub02PlanTasks, owners, and risks
  3. People03Review planPeople adjust and approve
  4. Agentic Hub04DelegateRoute each task with full context
  5. Specialist agents05ExecuteParallel where possibleEscalates to people when unsure
  6. Agentic Hub06CheckGuardrails and consistency
  7. People07ApproveConsequential work
  8. Tools and channels08ActPublish, send, or update
  9. Agentic Hub09LearnMeasure and log results
  10. Results inform the next plan, back at step 02.
An orchestration run across four lanes. Dashed lines show escalation to people and the learning loop back to planning.

The autonomy ladder

Five levels of agent independence, and the evidence needed to climb them.

Level 0ManualPeople do the workFull human controlLevel 1AssistAgent drafts andresearchesPeople choose what touseLevel 2Approve to actAgent preparesactionsPeople approve eachoneLevel 3Act and notifyAgent acts withinlimitsPeople can reverseLevel 4AutonomousAgent acts withinguardrailsPeople review samplesHard guardrails enforced in code from level 2 upwardPromote one level at a time, with evidenceStep down when the failure budget is spent
  1. Level 0ManualPeople do the workFull human control
  2. Level 1AssistAgent drafts and researchesPeople choose what to use
  3. Level 2Approve to actAgent prepares actionsPeople approve each one; guardrails in code
  4. Level 3Act and notifyAgent acts within limitsPeople can reverse; guardrails in code
  5. Level 4AutonomousAgent acts within guardrailsPeople review samples; guardrails in code
  6. Promote one level at a time, with evidence. Step down when the failure budget is spent.
The autonomy ladder. Higher levels require more evidence and hard guardrails enforced outside the model.

Evaluation gates

The stages an agent passes before it touches live work, and the criteria at each gate.

Stage 1Offline evalsGolden dataset andrubricStage 2Red teamDeliberate failurehuntingStage 3Shadow modeReal inputs, no realactionsStage 4Limited rolloutAssist or approve toactStage 5ProductionMonitoring and failurebudgetMeets qualitybar on test setGateSevere failuresfixedGateMatches humanbaselineGateWithin failurebudgetGateNew failures in production become new test casesWork only moves right when a gate's criteria are met
  1. Stage 1Offline evalsGolden dataset and rubric
  2. Gate: meets quality bar on test set
  3. Stage 2Red teamDeliberate failure hunting
  4. Gate: severe failures fixed
  5. Stage 3Shadow modeReal inputs, no real actions
  6. Gate: matches human baseline
  7. Stage 4Limited rolloutAssist or approve to act
  8. Gate: within failure budget
  9. Stage 5ProductionMonitoring and failure budget
  10. New failures in production become new test cases at stage 1.
Evaluation gates. Work moves right only when a gate's criteria are met; production failures become new test cases.

Signal data flow

How data moves from sources to agents and channels, and how outcomes flow back.

SourcesWebsiteProduct usageCRMEmail and adsEventsSupportSignal data layerIdentityRecords linked conservativelyConsentPreferences and opt-outsIntentMeaningful signal patternsEvent streamOne timeline, people andagentsOnly eligibledata passesAgentic HubDecides what to dowith each signal,within guardrailsAgents and channelsSales alertsBriefs for accountownersLifecycleTimely, consentedmessagesPaid mediaAudiences and budgetsContentTopics buyers askaboutOutcomesMeetings,pipeline,retentionAttribution and results feed back into the data layer
  1. Sources
    • Website
    • Product usage
    • CRM
    • Email and ads
    • Events
    • Support
  2. Signal data layer
    • IdentityLinked conservatively
    • ConsentPreferences, opt-outs
    • IntentMeaningful patterns
    • Event streamOne timeline
  3. Consent gate: only eligible data passes
  4. HubAgentic HubDecides what to do with each signal, within guardrails
  5. Agents and channels
    • Sales alerts
    • Lifecycle
    • Paid media
    • Content
  6. ResultOutcomesMeetings, pipeline, retention
  7. Attribution and results feed back into the data layer.
Signal data flow. A consent gate sits between the data layer and the hub, and results feed back into the data layer.

Four ideas behind every diagram

People set goals and limits

Agents pursue goals that people define, inside guardrails that people write and code enforces.

One hub coordinates

Plans, shared facts, routing, and approvals live in one place, so many agents behave like one system.

Everything is recorded

Every run, decision, and outcome is logged, which is what makes evaluation, audit, and learning possible.

Autonomy is earned

Agents gain independence one level at a time, with evidence, and step back down when the evidence changes.

See the architecture in the product

Marketix is built on these diagrams: specialist agents, governed connectors, and a hub that plans, routes, and records every run. Sign up for early access.