AI patterns that ship.

Start with the smallest useful shape: a direct agent, a grounded answer, a workflow, or a review gate. Add orchestration only when the product risk asks for it.

Source-grounded answer RAG becomes a visible pipeline, not a hidden prompt.
Use Case 01

RAG & Source-Grounded Q&A

Build retrieval-augmented generation systems where the agent searches, reads, cites, and validates sources — all with typed output schemas.

  • Search and read are explicit tool calls.
  • The answer carries citations and confidence.
  • Unsupported claims are rejected before publish.
Human-in-the-loop The write path only opens after approval.
Use Case 02

Human-in-the-Loop Review

When proposed changes should not be applied until a human approves them. The proposal agent drafts; the workflow owns the gate.

  • No mutation before approval.
  • Approve, revise, and reject are separate paths.
  • Every decision leaves an OTel trace and log entry.
Typed routing Route on structured fields, not prose.
Use Case 03

Triage & Routing

Classify incoming requests, assign priority and ownership, and route to the right downstream handler — all with narrow, typed outputs.

  • Classification produces narrow JSON fields.
  • Routing policy consumes fields, not prose.
  • Escalation is just another typed route.
Fan-out / fan-in Specialists run independently, then merge.
Use Case 04

Parallel Agents

Fan-out to multiple specialized agents for code review, risk analysis, test generation, and documentation — then synthesize results into a single coherent output.

  • Each branch has its own instructions and tools.
  • The workflow owns fan-out and fan-in.
  • Synthesis returns one typed result.
Inspectable reasoning Correctness work is split into named steps.
Use Case 05

Plan, Reason,Reflect, Judge

When correctness matters more than raw speed, use a structured workflow to plan, retrieve, reason, reflect, judge, and publish.

  • Each phase is named and traceable.
  • Reflection and judging are explicit steps.
  • Publishing happens only after readiness checks.
Living knowledge loop Sources feed answers; audits feed maintenance.

The loop is intentional: every generated answer can become evidence for what should be updated next.

Use Case 06

Living Wiki

A self-updating knowledge base that combines direct Q&A, ingest workflows, decision memos, architecture reviews, and continuous audit — all with human review gates.

  • Ingest turns sources into grounded references.
  • Q&A uses citations instead of memory guesses.
  • Audit work feeds reviewed maintenance.

Add memory and context.

Learn how agents remember across sessions, runs, and users without turning memory into hidden business truth.

Memory & Context → Architecture Reference