Webex AI Agent

Each autonomous agent carries up to ten actions. MCP client actions connect outside tools and authenticate with OAuth2 client credentials, an API key or custom headers, and each tool run is logged as a success or failure.

Fulfillment actions run Webex Connect Flow Builder flows, the route by which an agent reaches back office and CRM systems through Cisco's integration layer, and custom transfer actions route to another AI agent, a person, voicemail or a hunt group. Input entities the action needs are defined in a table or a JSON schema and filled from the conversation. MCP actions cannot be edited once created, so a change means a new action.

Sourcehelp.webex.com/en-us/article/ncs9r37/webex-ai-agent-studio-administration-guideread 2026-10-11

An autonomous agent works toward a goal written as instructions, deciding from the conversation which of its actions to run and collecting the inputs each needs, while multistep back office work runs in Webex Connect Flow Builder flows the agent calls, with a fulfillment timeout set between 10 and 30 seconds. Custom transfers pass a conversation from one AI agent to another or to a person, announced or silent, so a team can split work across specialist agents. Scripted agents follow defined rules for structured tasks alongside them, and a tenant can hold up to 100 agents of both kinds.

Sourcehelp.webex.com/en-us/article/ncs9r37/webex-ai-agent-studio-administration-guideread 2026-10-11

Knowledge bases take files such as PDFs, Word documents, Excel workbooks and CSV files, written articles and website crawls, up to 2GB and 100 files per knowledge base, 10MB per file and 300 pages per PDF, and one knowledge base can serve several agents.

A crawl reaches up to 250 pages at a depth of up to two links, across up to 10 URL patterns and 10 subdomains, and spreadsheets can hold up to 3,000,000 characters. Web sources show a syncing status while content is extracted, and with content approval on, a sync waits as pending approval until someone approves or declines the changes. Payment card, personal and health data are meant to stay out of knowledge bases.

Sourcehelp.webex.com/en-us/article/ncs9r37/webex-ai-agent-studio-administration-guideread 2026-10-11

Agent handover to a human is on by default for every autonomous agent, and custom transfers send callers to a person, a queue or voicemail. Built in guardrails keep the agent from unethical or harmful replies, the instruction optimizer tones down abusive or overly strong wording for the team to accept or discard, and knowledge syncs can be held for a person to approve. EU deployments show an AI transparency notice by default, and previews can show it too. No person approves an individual action before it runs.

Sourcehelp.webex.com/en-us/article/ncs9r37/webex-ai-agent-studio-administration-guideread 2026-10-11

Five administrator roles, full, contact center service, provisioning, read only and supervisor, govern AI Agent Studio through Control Hub or Webex Connect, and custom roles can carry narrower permissions such as viewing the audit log.

Transcript access is off by default until a full administrator grants decryption through the enterprise profile, tenant encryption keys sit in a key management service, no Cisco employee accesses raw audio or text during inference, and recording playback switches off when payment card data is detected.

Webex as a whole holds ISO/IEC 27001 certification and supports HIPAA and GDPR use, but no certification is tied to AI Agent itself, and MCP actions do not support group level access control.

Sourcehelp.webex.com/en-us/article/ncs9r37/webex-ai-agent-studio-administration-guideread 2026-10-11

Session lists and session details show each conversation an agent handled, filtered by test sessions, handovers, errors or downvoted replies, and MCP tool runs record success or failure.

Analytics count sessions, sessions resolved with no human involved, handovers, messages and users, each with daily averages, and custom reports cover AI agent call records and outcome distribution, with scripted agents adding views for training, responses and curation.

Each publish saves a version, and change logs record who changed what, where and when, searchable by admins, AI agent developers and custom roles with audit log access.

Sourcehelp.webex.com/en-us/article/ncs9r37/webex-ai-agent-studio-administration-guideread 2026-10-11

Within a conversation the agent fills input entities and keeps them for its actions, and global variables pass values into reports. Each session starts fresh. The studio keeps no memory that carries what a customer said in one conversation into the next, and transcripts are purged automatically after 90 days.

Sourcehelp.webex.com/en-us/article/ncs9r37/webex-ai-agent-studio-administration-guideread 2026-10-11

Contact center data is stored in one of eight regions, the United States, the United Kingdom, the European Union, Canada, Japan, Australia, Singapore and India, and Webex AI Pro US and Europe engines serve customers in the US and the EU. GPU based model and speech processing can run outside the home region, Azure OpenAI runs globally for several regions, and EU inference stays in the EU even during failover. Cisco's residency tables disagree with each other in places, so exact processing locations outside the EU carry some uncertainty.

Sourcehelp.webex.com/en-us/article/9qrsly/data-security,-privacy,-and-residency-in-webex-contact-centerread 2026-10-11

A new AI agent starts from scratch or from a predefined template, for scripted and autonomous agents alike, and a tenant holds up to 100 agents across both types. Each agent's details export as JSON. Cisco does not name the templates outside the studio itself.

Sourcehelp.webex.com/en-us/article/ncs9r37/webex-ai-agent-studio-administration-guideread 2026-10-11

Calls and digital chats reach the agent through the contact center's own entry points and routing, in Webex Contact Center, Webex Contact Center Enterprise, Unified CCE, Packaged CCE and, in an alpha release, Unified CCX. Digital agents come with the voice subscription order, and Webex Connect Flow Builder hands chats to an agent through its Virtual Agent V2 activity.

The agent answers before a customer reaches a person and hands back into the same queues. On voice, callers can interrupt the agent mid reply, and a session closes after three prompts in a row go unanswered. Cisco names voice and chat as the channels, without SMS or WhatsApp.

Sourcehelp.webex.com/en-us/article/ncs9r37/webex-ai-agent-studio-administration-guideread 2026-10-11

Each agent runs on an AI engine the team picks from the Webex AI Pro family, with standard, US and Europe variants in 1.0 and 2.0 versions and the 1.0 engines due for deprecation. Pro 2.0 handles multilingual contact centers, while the US and Europe engines are English only, and scripted agents run on Webex AI Agent Pro engines with Swiftmatch. The engines are Cisco's own packaging over models that include Azure OpenAI, and a team cannot pick a model provider directly.

Sourcehelp.webex.com/en-us/article/ncs9r37/webex-ai-agent-studio-administration-guideread 2026-10-11

Webex for Developers carries Webex Contact Center APIs and a web SDK for contact center applications. Agents themselves are built, tested and published inside the studio, with each agent's details exportable as JSON, and the MCP support in the studio points outward to other tools.

Sourcedeveloper.webex.com/webex-contact-center/docs/webex-contact-centerread 2026-10-11

Agents can be previewed at creation, while being edited and after deployment, through a preview widget, a chat preview or a voice preview, with voice preview available through Control Hub. The Sessions page separates test sessions from live ones and filters handovers, errors and downvoted replies, and scripted agents have their own test and curation views in analytics. Autonomous agents have no saved test cases with pass or fail results.

Sourcehelp.webex.com/en-us/article/ncs9r37/webex-ai-agent-studio-administration-guideread 2026-10-11

Agents act through MCP tools, Webex Connect flows and transfers, and website crawls feed knowledge bases when content is added. No hosted browser, remote desktop or screen control is part of the studio. Voice agents collect keypad entries from callers, with digit timeouts and lengths of up to 32 characters, but never drive another system's screen.

Sourcehelp.webex.com/en-us/article/ncs9r37/webex-ai-agent-studio-administration-guideread 2026-10-11