--- canonical: https://docs.jentic.com/concepts/business-agents/ representation: markdown-alternate # Canonical page: https://docs.jentic.com/concepts/business-agents/ # This is the Markdown alternate of the page above. title: "Business Agents Explained — Jentic Concepts" description: >- Understand business agents: AI systems that plan and execute real-world tasks by calling external APIs and workflows securely, observably, under your controls. --- # Business Agents **Business agents** are AI systems that plan and execute real‑world tasks by calling external **APIs** and **workflows**—securely, observably, and under your organization’s controls. > TL;DR: The world runs on APIs. Business agents turn API catalogs into *capabilities* and execute them reliably with proper auth, permissions, and monitoring. --- ## Why now? * **APIs everywhere** → Your business logic already lives behind APIs and SaaS tools. * **LLMs are planners** → Given tool descriptions and guardrails, they can choose the right capability. * **Standards matured** → **OpenAPI**, **Arazzo** (workflows), and **MCP** (tooling transport) enable interoperability. * **Enterprise needs** → Security, audit, observability, and repeatability are non‑negotiable. --- ## Core principles * **Capabilities, not code** — Agents consume *descriptions* of operations and workflows, not bespoke glue code. * **Just‑in‑Time Tooling (JITT)** — Load tools on demand to reduce hallucination and keep the agent’s context clean. * **Auth‑aware planning** — Agents only plan with what they’re **authorized** to use (AuthN + AuthZ). * **Workflows first** — Deterministic recipes deliver reliability, speed, and cost efficiency. * **Open by design** — Built on **OpenAPI**, **Arazzo**, and **MCP** for portability and future‑proofing. * **Governed & observable** — Centralized credentials, permissions, and fleet telemetry. --- ## A mental model ``` [User / System] ── goal ─▶ [Agent Planner] ──(needs tools)──▶ [Capability Catalog] │ (OpenAPI + Arazzo) │ plan + steps ▼ [Agent Executor] ──calls──▶ [APIs / Workflows] │ ▲ │ telemetry │ auth ▼ │ [Observability & Policy] ◀── [Central Auth/Perms] ``` --- ## How Jentic fits Jentic is **one platform** for business‑agent capabilities: * **Capability Catalog** — 10,000+ APIs and agent‑ready **workflows**; all **open‑source** and standards‑based. * **Secure by default** — **Centralized credentials**, credential injection, unified permissions. * **JITT tooling** — Load only what the agent needs, when it needs it. * **Observability** — See your **agent fleet**, calls, and data flows for governance. * **Everywhere** — Use with Cursor, Windsurf, Claude Desktop (via **MCP**), cloud agent platforms, or your bespoke agents. **Get started** * Onboard fast: [Quickstart](../getting-started/quickstart.md) * Use Arazzo workflows: [Arazzo Runner](../getting-started/arazzo-runner.md) * Connect IDEs/clients: [MCP overview](../guides/mcp/remote-mcp.md) --- ## Capabilities vs. Workflows * **Capabilities** (from **OpenAPI**): atomic, typed operations (e.g., “Send a Discord DM”). * **Workflows** (from **Arazzo**): deterministic multi‑step recipes with inputs/outputs (e.g., “Create Zendesk ticket, then notify Slack with ticket URL”). Why workflows win for production: * Fewer LLM decisions at runtime → **higher reliability** * Reusable, testable, reviewable → **better governance** * Cached patterns → **lower latency and cost** --- ## Security & governance * **Credential Injection** — Credentials stored centrally and injected at execution time; they never reach the agent, its prompts, or client code. * **Unified Permissions** — Scope what each agent can call; enforce org policy across agent fleets. * **Audit & Observability** — End‑to‑end tracing of tools, workflows, and data movement. See: **[Credentials & keys](../guides/credentials/jira-guide.md)** --- ## Reliability by design (JITT) Preloading dozens of tools in context invites confusion. With **Just‑in‑Time Tooling**, the agent: 1. Plans with known/authorized capabilities 2. **Loads** the specific tools it needs 3. Executes with validated inputs 4. Emits telemetry for continuous improvement Result: **Fewer hallucinations, better parameterization, stable outcomes.** --- ## Where business agents live * **Prebuilt agents:** Cursor, Windsurf, Claude Desktop (connect via **MCP**) * **Cloud stacks:** AWS Bedrock, Microsoft Copilot Studio, Google Vertex AI * **Custom agents:** Your own framework, driving your self-hosted [Jentic One](../reference/jentic-one/index.md) instance via the `jentic` CLI or MCP --- ## Example outcomes * **Support Ops:** auto‑triage emails → create Zendesk ticket → summarize to Slack. * **RevOps:** enrich a lead → add to CRM → send intro email via provider. * **FinOps:** pull invoices → detect anomalies → file to data warehouse and alert. * **IT Ops:** rotate a secret → update dependent services → notify on-call channel. --- ## Anti‑patterns to avoid * Stuffing secrets into prompts or agent memory * Hard‑coding API glue for each tool/vendor * Preloading massive tool lists into every LLM call * Relying on ad‑hoc, uncontrolled “scripts” without auditability --- ## FAQ **Is Jentic required to build a business agent?** No—but Jentic eliminates a ton of undifferentiated heavy lifting (auth, cataloging, workflows, observability) and aligns everything to open standards. **Can I bring my own private APIs?** Yes. Jentic One is self-hosted — you add private APIs and workflows to your own instance and control who/what can access them. **Does this lock me in?** Jentic is **open‑standards–first** (OpenAPI, Arazzo, MCP). Your capabilities remain portable. --- ## Next steps * Try it end‑to‑end → **[Quickstart](../getting-started/quickstart.md)** * Use it from IDEs → **[MCP overview](../guides/mcp/remote-mcp.md)** — hosted directory search + your instance as a custom connector * Questions? → **[Contact us](../community/contact-us.md)** > Want to help shape the catalog? See **[Contributing](../community/contributing.md)**. --- ### Appendix: term glossary * **Capability** — a typed, documented operation (typically from OpenAPI). * **Workflow** — a deterministic multi‑step recipe (Arazzo). * **MCP** — Model Context Protocol; transports tools to agent clients. * **JITT** — Just‑in‑Time Tooling; load tools only when needed. * **AuthN/AuthZ** — Authentication/Authorization.