Infobip

Three named tool types with authenticated action, configured per agent. In the Tools section a customer adds a Component, an MCP server or an Exchange application. Tools are documented as letting the agent call APIs, execute backend logic, retrieve external data and trigger automated workflows. Components encapsulate reusable logic and can orchestrate multiple API calls or backend processes for complex integrations.

MCP servers are configured with a name, endpoint URL, HTTP headers for authentication tokens, and an Allowed functions field naming which actions the server may take when called, so the connection is authenticated and scoped rather than open. Exchange applications are selected from a published integrations catalog. Underneath sits the messaging estate the agent acts through: more than fifteen natively integrated channels each with its own API, plus the modular CPaaS X stack.

Sourceinfobip.com/docs/ai-agents/ai-agents/configure-ai-agentread 2026-09-05

Named multi-agent delegation plus a rule-based flow layer, both documented as product surfaces. An AI agent can be configured as an orchestrator that coordinates other agents: the customer adds subagents from a list, and the orchestrator delegates tasks to specialized agents, combines their responses and controls conversation routing, with the agents table showing a connected-subagents count for orchestrators.

Each agent's Description field is documented as what other agents read to decide when to call it, which is a published delegation contract rather than an implied one. Beneath that, Automation Studio builds journey flows and chatbots that route on rules, triggers and Customer Profiles data, and the Agent connector element drops an AI agent into a chatbot so deterministic flows and autonomous agents combine in one path. Agents themselves execute multi-step tasks, calling tools in sequence to reach a goal.

Sourceinfobip.com/docs/ai-agents/ai-agents/configure-ai-agentread 2026-09-05

A maintained retrieval structure published as its own product line. Knowledge agents are documented as specialized agents for information retrieval from documents, URLs and files, purpose-built for FAQ-style interactions, and they sit alongside a Knowledge Base shared component available across AgentOS.

The corpus is maintained rather than assembled per request: Quality center carries a dedicated Knowledge agents section that validates responses against expected answers and tracks quality over time, which only makes sense against a persistent body of content the customer curates. General-purpose AI agents reach the same material through tools and components. The Customer Profiles data platform is a database rather than a grounding corpus.

Sourceinfobip.com/docs/ai-agents/knowledge-agents/overviewread 2026-09-05

A configuration surface the operator sets, plus a documented escape path to a person. Guardrails are an explicit field in the agent's Advanced settings with their own configuration page: the customer selects content filters that detect and block harmful content in agent conversations, controlling what the agent can and cannot discuss. Release control sits with the customer too, since a configured agent stays a draft until a named version is created and explicitly activated for production, with earlier versions available to roll back to.

Scope on outbound action is set per tool, with each MCP server carrying an Allowed functions list naming what it may do when called. When automation reaches its limit the conversation transfers to Inbox with the complete history so the human agent does not restart the customer, and that human works with an Inbox copilot. No pre send approval step on an individual agent action is documented, so oversight comes from filters, scope and escalation rather than sign off on each action.

Sourceinfobip.com/docs/ai-agents/ai-agents/configure-guardrailsread 2026-09-05

Attestation and access are both documented. Infobip's own Information Security Terms state it is regularly audited against ISO/IEC 27001, 27017, 27018 and 22301 and the ISAE 3000 standard, assessed annually, with certificates available for download at infobip.com/certificates and a full or summary copy of current reports provided on written request, which is an obtainable report rather than an asserted badge.

A dedicated HIPAA attestation under SSAE 18 is published in Infobip's own newsroom. The legal tier is public rather than gated: sub-processor list, data transfer agreement, data retention notice and privacy notice all publish at named URLs.

On access, the docs cover single sign on through any SAML 2.0 identity provider with SP-initiated flow and identity-provider groups mapped to custom roles and re-synchronized on every login, role-based access control with predefined and custom roles governing per-feature permissions, two-factor authentication configurable as remembered or on each login, password validity and maximum login attempt policies, IP safelisting, and separate API access permissions. Infobip's identity, authentication and fraud prevention products are things it sells to its customers, not controls over the customer's own tenant.

Sourceinfobip.com/docs/essentials/manage-my-account/account-settingsread 2026-09-05

Per-run visibility plus change control, both named in the docs. Insights and Analytics is one of the five AgentOS products and is documented as monitoring AI agent performance, tool usage and session tracking through dedicated dashboards, with every interaction across all five products feeding it and the output covering performance metrics, root causes and AI-driven recommendations. Tool usage and session tracking is a record of what an individual agent did, not only an aggregate.

From inside an agent, Go to Session history opens previous sessions with their associated analytics, and View as code renders the running configuration. Change is auditable separately: agents autosave up to ten draft versions, named versions are created and activated deliberately, and version history records what was live when. No immutable audit log with actor and timestamp for configuration changes is documented, so what exists is execution visibility plus version control rather than a compliance grade trail.

Sourceinfobip.com/docs/insights-and-analyticsread 2026-09-05

A customer data store is documented and an agent memory layer is not, and these are different things. Customer Profiles is AgentOS's customer data platform, holding person and company profiles, attributes, segments and behavioral events, with interaction history available to AI agents so they can personalize responses and make context-aware decisions across sessions.

Agents also carry session state within a conversation, and past sessions are viewable through Session history. A database is not a memory layer whoever writes it, and Customer Profiles is published as exactly that, a store the customer loads and queries rather than state the agent forms and retains. There is no agent memory product, no documented scope, lifetime or deletion control for retained agent state, and no memory entry in the AgentOS product set.

Sourceinfobip.com/docs/customer-profilesread 2026-09-05

Multi-region placement is documented by name, customer-controlled selection of it is not. The AI Agents documentation states outright that AI agents are available only in the following data centers: Europe (EU1), Sweden (SE1), Germany (FR4), India (MUNM) and United States (NY2), so an account sits in a named region across three continents and the buyer can tell which.

Around that Infobip publishes the full legal tier rather than gating it: a sub-processor list, a data transfer agreement, a data retention notice and a privacy notice, each at its own public URL, plus an infrastructure overview and network connectivity options in the docs. No self service or documented mechanism for the customer to choose or move their region is published, placement appears to be set at account provisioning, and no on premise, private cloud or single tenant deployment is offered.

Sourceinfobip.com/docs/ai-agents/ai-agents/overviewread 2026-09-05

Adoptable units the buyer picks from a list rather than assembles. Agent instructions can be populated from predefined templates named in the docs as Appointment booking agent, Order status assistant and Sales representative, and Automation Studio carries its own predefined and custom templates to speed up journey and chatbot creation.

Agents are portable as assets: an existing agent can be duplicated, exported to a.export file and imported into another workspace, so a configured agent moves as a unit. Components are reusable logic blocks selected from a list and shared across agents. Above all of it, AgentOS itself ships as five named products a customer switches on rather than builds. The template set is small, and there is no browsable marketplace of ready made agents.

Sourceinfobip.com/docs/ai-agents/ai-agents/configure-ai-agentread 2026-09-05

A very wide channel surface, with three documented ways for work to start. Channels: WhatsApp, SMS, RCS, MMS, email, voice, live chat, video, in-app messaging, mobile push, Viber, Messenger, Instagram, Telegram, Apple Messages for Business, Kakao, Line and Zalo, each published with its own API page, and all of them flowing through the same system.

Initiation is documented three ways: inbound, where a customer reaches out on any supported channel; event-driven, where Automation Studio journey flows are triggered by customer behavior or profile changes; and scheduled, where journeys start on dates. Voice is first-class rather than bolted on, with a Transfer call to AI agent element placing an agent inside an IVR flow, and Broadcast handles outbound at volume.

Sourceinfobip.com/docs/automation-studioread 2026-09-05

The customer picks the model, and it is a field in the product rather than an inference. Under Advanced settings when configuring an AI agent, Model is documented as: select the LLM model that best matches your requirements for reasoning capability, latency and cost.

That is per-agent selection by the buyer against stated trade-offs, and it sits alongside Guardrails and structured output in the same settings panel. The Instructions field is described as acting as the system prompt for the underlying language model, confirming the customer configures the model layer directly. Third party agents reaching Infobip through MCP is interoperability, not model choice.

Sourceinfobip.com/docs/ai-agents/ai-agents/configure-ai-agentread 2026-09-05

An open, fully public developer surface, which is unusual. Infobip publishes an API reference, SDKs, release notes, an integrations catalog and a public roadmap, all reachable without a login, plus Postman collections, a GitHub organization and a developer Discord. Every channel is exposed as its own REST API and the modular CPaaS X stack is sold specifically for platforms building on top.

On the agent side extensibility runs both directions: MCP servers let an Infobip agent call outside systems with scoped allowed functions, and an agent's configuration can be viewed as code, exported to a portable file and imported elsewhere, with structured output constrained by a customer-defined JSON schema. Components let a developer encapsulate backend logic the agent then calls, and Exchange applications add partner integrations.

Sourceinfobip.com/docs/apiread 2026-09-05

A dedicated evaluation product with a readable comparable result. AgentOS ships Quality center, documented as tools to monitor, evaluate and improve agent response quality, where the customer creates test groups, defines test cases, runs tests and reviews results to confirm agents behave as expected.

For Knowledge agents it is stated as validating responses against expected answers and tracking quality over time, which is a scored comparison rather than a bare pass gate. Separately, the build flow carries a pre-publish step, Test an AI agent, described as validating agent responses and behavior before publishing, and agent versioning lets a team activate a specific version and roll back a bad one. The artifact under test is the customer's own agent.

Sourceinfobip.com/docs/ai-agents/quality-center/overviewread 2026-09-05

No browser control, hosted session, virtual desktop or interface automation appears anywhere in the AgentOS documentation, and the route out is stated positively rather than as a silence.

An Infobip agent reaches the world through exactly three documented tool types, all programmatic: Components running backend logic, MCP servers called at an endpoint URL with scoped allowed functions, and Exchange applications from the integrations catalog. Its output side is the messaging estate, more than fifteen channels each with its own API. Nothing in the agent configuration surface offers a browser, a screen or a pointer.

Sourceinfobip.com/docs/ai-agents/ai-agents/configure-ai-agentread 2026-09-05