What are Agent Stacks?

An Agent Stack is Atender’s AI customer service team — a virtual agent that handles inbound conversations from end to end. To customers it looks like a single helpful person with a name and an avatar. Under the hood it’s a layered system: an orchestrator that reads every message and dispatches it to the right expert, and a set of specialist agents that each have deep knowledge of one area.

A stack can run on whichever channels you enable — email, web chat, SMS, voice, WhatsApp, Messenger and Instagram — and conversations it’s working on appear in the inbox tab that lists Agent Stack conversations — it reads Agents, or the name of your default stack — in Conversations so your team always has visibility into what the AI is doing.

Name it whatever you want

You’re not stuck calling it “the AI.” Each stack has its own name, description and avatar, and the name is what visitors see as the display name when the stack replies — Sophie, James, Alia, Atender Bot. The default stack also lends its name to the inbox tab that lists AI-handled conversations; with no default stack, that tab reads Agents.

What’s inside a stack

The stack isn’t one prompt. It’s two layers working together:

• Orchestrator — Reads each customer message, classifies it against this stack’s specialists, and dispatches to the right one with full conversation context. It also owns the handover tool, handover_to_human; a specialist that cannot resolve a request can ask for an escalation with its own request_handover tool, which the platform then runs through the same handover flow
• Specialist agents — The experts. Each one has its own scope, instructions, knowledge access, and capabilities

The more specialists you add, the more capable the stack becomes. A stack with one generalist will never match a stack with five focused experts — each can have its own scope, its own instructions, and only the knowledge and capabilities relevant to its area. See Anatomy of an Agent Stack for the full architecture, and What is a specialist agent? for what makes a good specialist.

Knowledge sources

Agent Stacks can pull from several kinds of knowledge:

• Knowledge Base — your customer-facing help articles. The AI can quote and link to these.
• Knowledge Packs — platform-owned collections of reference material (the Library) that Atender grants to your workspace. Access is per specialist, not per stack: an admin assigns granted packs to a specialist on that specialist’s Knowledge sub-tab, and only the packs assigned to it take part in its retrieval.
• Handbook — your internal procedures and policies. The AI uses these to guide its behavior but never quotes them directly to customers.
• Code bases — GitHub repositories connected by admins in Settings. Specialists can be granted read-only code-base tools when they need product implementation context. Code bases are a fallback source: the agent should use them only when customer-facing Knowledge Base content and internal Handbook content do not answer the question. Even then, it should answer in product terms and must not quote code, file paths, or internal names to customers.

Capabilities — agents that can act

Without capabilities, an Agent Stack can only talk. With capabilities, it can act — look up an order, process a cancellation, check delivery status, verify account ownership, or call any external API you connect. Each specialist gets only the capabilities it needs, so a billing agent can issue refunds while the product agent can only look things up. On voice and phone channels, public capabilities are available without proof; Read and Act capabilities can become available after verify_caller succeeds when caller verification is enabled for the tenant; Transact capabilities are not callable on voice and route to humans. See current tier and verification-posture rules for the channel-specific caveat.

Handing off to humans

The orchestrator holds the handover tool, and a specialist that hits its limits can request an escalation the platform runs through the same flow. When the customer asks for a person and has said what they need, when the request is out of scope, or when they stay hostile after a couple of attempts, it calls that tool and the platform queues the conversation for a team. You control how this happens via the Handover tab — required information the AI must collect before handing off, the team that picks up, and what happens outside opening hours. See Handover to humans.

Testing and tuning

Every stack has a built-in test sandbox. You can chat as a customer, see how the orchestrator routes, watch which specialist responds, and flag any reply that missed the mark. Flagged responses feed the Tuning flow — Atender analyzes what went wrong, proposes a concrete change to one agent’s instructions or its responsibility, and you apply it or undo it with a click. Separately, Self-Learning works from the feedback your team leaves on real conversations in the inbox, proposing improvements based on patterns across many of them.

Where to start

• Build your first stack: Create an Agent Stack
• Understand the architecture: Anatomy of an Agent Stack
• Add expertise: Add a specialist agent
• Tune the personality: Set the stack personality
• Hand off correctly: Configure handover