4 Agentic Workflow Patterns

The building blocks for every automated workflow. Watch each pattern animate step-by-step with AnyCompany Finance examples.

๐Ÿ“– Day 3 Reference โšก Animated ๐Ÿฆ Finance Examples
Pattern 1

๐Ÿ”— Prompt Chaining

Sequential steps โ€” the output of one becomes the input of the next. Trades speed for accuracy. Each step can be tested and retried independently.

๐Ÿ“„ExtractRead invoice PDF
๐Ÿ”ValidateMatch vs PO
โš–๏ธDecideApprove / Escalate
๐Ÿ“ŠReportGenerate summary
Press Play to watch an invoice flow through the chaining pipeline step by step.
Step 0 / 4

๐Ÿ’ช Strengths

  • Easy to understand and debug
  • Each step testable independently
  • Preserves context across steps
  • Supports rollback and retry

๐Ÿฆ Best For (Finance)

  • Invoice processing pipelines
  • Report generation workflows
  • Sequential approval processes
  • Data extraction โ†’ validation โ†’ output
๐Ÿ’ก
Day 2 connection: Chaining = the task decomposition technique you practiced. Each box is a prompt template. The chain connects them.
Pattern 2

โšก Parallelization

Multiple tasks run simultaneously on the same input. An aggregator combines the results into a unified answer. Faster and more thorough than sequential.

๐Ÿ“‹Merchant DataInput to analyze
๐Ÿ›ก๏ธRisk AnalystConservative view
๐Ÿ“ˆProduct LeadGrowth potential
โš–๏ธComplianceRegulatory check
๐ŸŽฏAggregatorCombine โ†’ GREEN/AMBER/RED
Press Play to watch 3 perspectives analyze the same merchant data simultaneously.
Step 0 / 4

๐Ÿ’ช Strengths

  • 3x faster than sequential
  • Multiple perspectives on same data
  • Each branch is independent
  • Aggregator resolves conflicts

๐Ÿฆ Best For (Finance)

  • Merchant risk assessment (Risk + Product + Compliance)
  • Due diligence checks (ID + Credit + Fraud)
  • Multi-perspective analysis
  • Voting / consensus decisions
๐Ÿ’ก
Day 2 connection: This is the multi-agent framing exercise โ€” same data, 3 personas. Now automated as parallel agents.
Pattern 3

๐Ÿ”€ Routing

Classify the input and direct it to the right processing path. Like a smart mailroom โ€” each document type gets its own optimized workflow.

๐Ÿ“จIncoming DocUnknown type
๐Ÿท๏ธClassifierWhat type is this?
๐Ÿ“„Invoice FlowExtract โ†’ Validate โ†’ Approve
๐ŸงพReceipt FlowLog โ†’ Categorize โ†’ File
๐Ÿ“‘Contract FlowReview โ†’ Flag โ†’ Alert
Press Play to watch a document get classified and routed to the right workflow.
Step 0 / 4

๐Ÿ’ช Strengths

  • Handles diverse input types automatically
  • Each path optimized for its type
  • Can chain with secondary routers
  • Scales to many categories

๐Ÿฆ Best For (Finance)

  • Document classification & processing
  • Support ticket triage
  • Payment type routing (card / QR / BNPL)
  • Regulatory filing by jurisdiction
๐Ÿ’ก
Day 2 connection: Routing = persona-based prompting. Different prompts for different document types, selected automatically.
Pattern 4

๐ŸŽฏ Orchestration

The most complex pattern. A central orchestrator dynamically manages the workflow, spawning subtasks based on conditions and coordinating human-in-the-loop decisions.

Orchestrator
๐Ÿ“ฅIntakeParse application
๐Ÿ“ŠAssess RiskScore & classify
๐ŸšฆDecision GateThreshold routing
โœ…<$10KAuto-approve
๐Ÿค$10Kโ€“$50KAI + Human
๐Ÿ‘ค>$50KHuman review
Press Play to watch a credit application flow through the orchestrator with dynamic decision gates.
Step 0 / 5

๐Ÿ’ช Strengths

  • Handles complex conditional workflows
  • Spawns subtasks dynamically
  • Human-in-the-loop at decision gates
  • Adapts to unexpected situations

๐Ÿฆ Best For (Finance)

  • Credit facility approvals
  • Insurance claims processing
  • Complex compliance workflows
  • Any process with conditional branching
๐Ÿ’ก
Day 2 connection: Orchestration = CoT reasoning + guardrails. The orchestrator reasons about what to do next; guardrails define the decision boundaries.
Chain Parallel Route

๐Ÿงฉ Combined: PayLater Application

Most real workflows combine 2-3 patterns. This AnyCompany PayLater application uses routing โ†’ parallelization โ†’ chaining in one pipeline.

๐Ÿ“‹ApplicationPayLater request
๐Ÿท๏ธClassifyNew / Existing / Renewal
๐ŸชชIdentityKYC verification
๐Ÿ’ณCredit BureauScore check
๐Ÿ”Fraud ScreenPattern analysis
๐Ÿ“ŠRisk ScoreCombine results
๐ŸšฆDecisionGREEN / AMBER / RED
Press Play to watch all 3 patterns work together in one PayLater application pipeline.
Step 0 / 6
๐Ÿงฉ
Key insight: Don't force everything into one pattern. Identify which pattern fits each PART of your process, then combine them. When you design your workflow in the afternoon exercise, label each step with its pattern.

๐Ÿ”— Chaining steps

  • Risk Score โ†’ Decision โ†’ Notification
  • Sequential, context-preserving

โšก Parallel steps

  • Identity + Credit + Fraud (simultaneous)
  • 3x faster than sequential

๐Ÿ”€ Routing step

  • Classify application type first
  • Different paths for new vs renewal

๐Ÿ“ Your workflow

  • Which patterns fit YOUR process?
  • Fill in the Agent Design Canvas