Every term from the workshop in plain English — agentic AI, Amazon Quick, and banking/finance shorthand. Searchable. Print it.
| Term | Meaning |
|---|---|
| Agent | A system that perceives input, reasons over multiple steps, and takes action — closing the loop a plain chatbot leaves open |
| Perceive → Reason → Act | The core loop every agent runs, regardless of complexity |
| Tool Use | An agent's ability to call an external function (API, database, calculator) and use the result |
| Chain-of-Thought (CoT) | Reasoning broken into visible intermediate steps rather than a single leap to an answer |
| Multimodal | Ability to process inputs beyond text — scanned documents, images, screenshots |
| Workflow Architecture | A fixed, deterministic sequence of steps with LLM reasoning inside each — predictable and auditable |
| Autonomous Architecture | The agent decides its own next step at each turn without a fixed script |
| Hybrid Architecture | A deterministic skeleton with autonomous reasoning inside individual steps — the most common finance pattern |
| Multi-Agent Architecture | Several specialized agents coordinated by an orchestrator |
| Chaining | Workflow pattern where each step depends on the prior step's output, run in fixed order |
| Parallelization | Workflow pattern where independent sub-tasks run simultaneously, then results are merged |
| Routing | Workflow pattern where a classifier step sends a request down the correct path |
| Orchestration | Workflow pattern where a coordinator agent spawns and manages sub-agents toward a larger goal |
| Space | A container holding documents and scoping an agent's knowledge — "the kitchen" |
| Chat Agent | A saved, reusable prompt configuration with persona, instructions, knowledge, actions, and suggested prompts — "the recipe" |
| Flow | A no-code, multi-step wrapper around one or more Chat Agents, schedulable and shareable — "the doorbell" |
| MCP Connector | Model Context Protocol integration giving an agent live tool access to external systems — "the oven" |
| Quick Research | An AI-led, multi-step investigation feature that produces a cited report |
| Pilot Canvas | Capstone deliverable — a structured one-pager for a proposed real agentic pilot, scored on autonomy/tools/guardrails/impact |
| Rule of Three | Heuristic: typed the same prompt 3 times → save it as an agent |
| Set-and-Forget Trap | Failure mode — an agent built once, never revisited as policy changes, quietly drifting out of compliance |
| Scope Creep Spiral | Failure mode — one agent slowly absorbing too many responsibilities until its behavior is unpredictable |
| Ghost Agent | Failure mode — an agent built once, shared, then silently abandoned |
| Audit-Defensibility | The finance-specific design pattern: named output sections, traceable numbers, flag-only behavior, cited sources |
| Flag-only behavior | An agent's guardrail that it can identify/classify an issue but never auto-resolve, approve, or clear it |
| Human-in-the-loop | A required checkpoint where a person must review or approve before an action is final |
| Break (reconciliation) | A line item where two records (e.g., GL and bank statement) don't match |
| Disposition memo/note | A drafted explanation of how an exception (break, NCR-style finding) should be classified and handled, for human sign-off |
| KYC / AML | Know Your Customer / Anti-Money Laundering — the due-diligence and monitoring controls referenced throughout finance agent guardrails |
| Maker-Checker | A control requiring the person who initiates an action and the person who approves it to be different individuals |