AI & automation

Agentic AI, MCP & custom tooling

Give AI agents the tools to act – data actions, A2A connections and MCP servers – specified, secured, audited and proven by one certified Genesys Cloud CX team.

AI that answers, acts and assists On the left a royal blue tile holds a virtual agent, drawn as a simple bot face, connected to three tool chips it calls to act: an automation, a data lookup and a record update. An arrow hands the conversation to the right-hand tile, where a human agent with a headset is assisted by a copilot suggestion bubble. A shared knowledge strip along the bottom feeds both the virtual agent and the copilot. AI & AUTOMATION Virtual agent Tools Copilot Knowledge

In summary

Genesys Cloud agentic AI moves automation from answering questions to taking actions – checking, booking, updating and resolving on the customer's behalf. QVCCS designs and builds the tools that make that safe: data action tools for agentic virtual agents, A2A connections to external AI agents, and bespoke Model Context Protocol (MCP) servers for approved AI clients. Architects, developers and testers from our own bench specify every tool in an interface contract, scope and authenticate it, log every action for audit and prove it under adversarial testing, so AI agents act across the enterprise without acting outside their remit.

Who works on this

  • Enterprise Solution Architect
  • Solution Architect
  • Senior Developer
  • Software Developer
  • Systems Integration Tester
  • Senior Platform Practice Lead (Genesys Cloud CX)
  • Tools built to purposeEach action an AI agent can take is a deliberately designed tool with clear inputs, validated outputs, masked data and named error cases.
  • Agents that collaborateA2A connections let a Genesys agentic virtual agent delegate specialist tasks to external AI agents while keeping the conversation.
  • MCP servers, built by usBespoke Model Context Protocol servers expose narrowly chosen Genesys Cloud and business capabilities to approved AI clients.
  • Scoped, audited, governedLeast-privilege credentials, action logging and layered guardrails make every automated step explainable to risk and compliance teams.
DiagramThe agentic AI stack around Genesys Cloud CX
  1. AI agents
    • Agentic agents
    • External agents
    • Copilots
  2. Protocols
    • Agent2Agent (A2A)
    • MCP
    • REST and events
  3. Tools built by QVCCS
    • Data actions
    • MCP servers
    • Functions and APIs
  4. Controls
    • OAuth scoping
    • Audit logging
    • Guardrails
  5. Systems of record
    • Genesys Cloud API
    • CRM and ERP
    • Bespoke systems

Agents reason at the top; QVCCS designs the tools, protocols and controls beneath them.

01

Why Genesys Cloud agentic AI needs engineered tools

Genesys Cloud agentic AI shifts the question from "what can the bot say?" to "what can the agent do?". An agentic virtual agent can reason across a customer's request, decide which actions to take and carry them out – checking eligibility, changing a delivery, issuing a refund within policy – then confirm the result in the same conversation. Genesys has launched AI agents with greater autonomy, and in October 2025 announced a partnership with, and investment in, Scaled Cognition on Large Action Models built for reliable action-taking. However capable the model, every action still runs through a tool someone has to design, build and secure.

That is where most of the risk and most of the value sits. A poorly scoped tool can expose more data than intended, act on the wrong record or fail silently. A well-designed tool does one job, validates its inputs, returns predictable results and leaves an audit trail. QVCCS treats tools as the engineering heart of agentic AI. In Define we capture an authority model – which tasks an AI agent may complete, on whose authority and where a person must approve – as functional and non-functional requirements, so security, operations and risk teams approve the scope before a line of code is written.

02

Tools, A2A and MCP on Genesys Cloud today

Inside Genesys Cloud CX, tools for agentic virtual agents are data actions, giving AI agents real-time access to your systems through integrations that are already permissioned and monitored. Genesys provides rich tool configuration: a name and description the agent uses to decide when to call it, pre-instructions such as authentication conditions, input variables, output validation in a guided builder or as Python expressions, variable masking that hides tokens and internal IDs from the agent, and named error cases with default instructions. QVCCS specifies each of these in an Interface Control Document before build, because the quality of the description directly shapes the agent's behaviour.

Open protocols are reshaping how AI agents connect. In September 2025 Genesys announced native interoperability with the Agent2Agent (A2A) protocol and the Model Context Protocol (MCP) for its Copilots and Virtual Agents. A2A is now documented: through the External A2A Server integration, still marked as coming soon, an agentic virtual agent acts as an A2A client, delegating specialist tasks to external agents on platforms such as ServiceNow or Salesforce while remaining the sole owner of the customer conversation and sharing only mapped fields. Genesys's documentation currently names A2A as the supported protocol for external agentic integrations, with further protocols planned, and does not yet document a native MCP capability, so we design MCP integrations to adopt it once Genesys makes it available.

Genesys has also extended its MCP reach through acquisition. On 30 June 2026 it announced the acquisition of Pinkfish, an agentic orchestration and workflow company, to expand Genesys Cloud AI with MCP-based tool integration and workflow automation. Genesys says Pinkfish brings more than 500 integrations supporting 25,000 MCP tools across CRM, ERP, IT, HR, order management and billing systems, so that agentic virtual agents and Copilots can securely reach data and act across the enterprise. Genesys expected Pinkfish capabilities through the AppFoundry by the end of July 2026 and native capabilities within Genesys Cloud by the end of January 2027, so we confirm current availability with Genesys for each design.

An AI agent is only as trustworthy as the tools you hand it – so we design every tool as if it will be audited, because it will.

QVCCS point of view

03

Authentication, scoping, audit and guardrails

Security for agentic AI starts with identity. Every tool, A2A connection and MCP server needs a clear answer to "on whose authority is this action taken?" – the customer's, an employee's or a service account's. Our architects design OAuth clients and roles with the minimum permissions each tool needs, separate read from write, and use divisions and separate development, test and production orgs so test agents never touch live data. Where an action depends on the customer being verified, the tool checks authentication state itself rather than trusting the conversation to have done so, and A2A field mappings share only what the external agent needs.

Audit is what makes autonomy acceptable to regulators and boards. We design logging so every tool call records which agent invoked it, for which conversation, with what inputs and what result, while keeping sensitive personal data out of logs. Guardrails are layered: Genesys built-in and custom guardrail rules, validation inside each tool, business rules such as refund thresholds that require a person, and monitoring that alerts engineers to unusual patterns. At Xperience 2026 Genesys introduced the AI Control Plane, providing centralised discovery, identity, policy and observability for agentic orchestration, and we assess how it complements your own controls.

Community MCP servers and open-source samples for the Genesys Cloud APIs are valuable for learning, but they are not automatically fit for a production contact centre. Before any third-party or community component is used, QVCCS reviews what it exposes, how it authenticates, how it is maintained and whether its scope can be narrowed. Often the right answer is a purpose-built server offering a handful of well-defined tools rather than a general gateway to every API. Each decision is recorded in an architecture decision record, so your security team can see exactly what an AI client is able to do and why.

04

Who engineers agentic tools, and what can go wrong

Agentic tooling crosses conversation design, integration, identity and operations, so we build a multidisciplinary team from our own bench with security at its centre. The Enterprise Solution Architect frames the authority model and how AI agents fit your wider estate; a Solution Architect produces the tool catalogue and an interface contract per tool; Senior Developers and Software Developers build data actions, functions, A2A configuration and MCP servers with CX as Code and version control under four-eyes peer review. Our Systems Integration Tester then tests decisions as well as code: agents face ambiguous requests, conflicting data and failing systems, and adversarial cases such as prompt-injection attempts, requests beyond remit, expired credentials and A2A partners that never reply. Pilots start narrow, and autonomy widens only as the audit evidence builds.

The pitfalls are mostly about tool design rather than models. A tool description that is vague or overlaps with another tool's leads the agent to call the wrong one, so we write descriptions as carefully as interface specifications and test them with real utterances. Tools that are too broad, such as a general "update customer" action, invite misuse; several narrow tools are easier to secure and audit. Write actions must be safe to retry, because a timeout followed by a second attempt should never issue two refunds. Irreversible steps need an explicit confirmation from the customer, and some need a person. Finally, every tool needs a defined failure message and a route to a human agent, so a broken back-end system produces a graceful handover rather than a confused conversation.

05

Running and evolving an agentic estate

Agentic AI is moving quickly, and your tooling has to keep pace without losing control. Release readiness reviews track Genesys changes such as agent model transitions, A2A availability, MCP support and the Navigator and Orchestrator capabilities Genesys expects to make generally available between late 2026 and April 2027. Each change gets a release impact assessment and a regression run of the scenario and adversarial packs before tools or autonomy are extended. Audit records are reviewed for actions that were refused, escalated or retried, which often reveal a tool that needs a sharper description or a tighter scope. QVCCS supports the data actions, A2A connections and MCP servers in line with your support model, as the specialist escalation layer behind your provider or under a direct SLA with monitoring and engineer callout.

What you get from QVCCS

  • Authority model and FRS/NFR for every action an agent may take
  • Tool catalogue with an Interface Control Document per tool
  • Data action tools with validation, masking and named error cases
  • A2A connection design and bespoke, narrowly scoped MCP servers
  • Audit logging, layered guardrails and human approval thresholds
  • Scenario and adversarial test packs traced to the RTM
  • Protocol release reviews, with support alongside your provider or direct from us

Methods & templates

Agentic AI, MCP & custom tooling: the templates behind each stage

Every QVCCS engagement follows our seven-stage delivery lifecycle, each stage closed by a quality gate. For Agentic AI, MCP & custom tooling, the work concentrates in Define, Design, Build, Prove and Run & evolve, and these are the techniques and standard templates we lean on there, each one traceable from requirement to design, build, test and support.

  1. 01DiscoverDiscovery sign-off
  2. 02DefineRequirements baseline
  3. 03DesignDesign authority review
  4. 04BuildBuild complete
  5. 05ProveGo / no-go readiness
  6. 06TransitionOperational acceptance
  7. 07Run & evolveService reviews
  • 2 · Define

    Functional & Non-Functional Requirements Specification

    Captures the authority model, approval thresholds, audit, latency and security requirements every agentic tool must satisfy.

  • 3 · Design

    Interface Control Document (ICD)

    Specifies each tool, A2A connection and MCP capability: purpose, inputs, output validation, masked fields, errors and identity used.

  • 3 · Design

    Architecture decision records

    Records why each capability is exposed through a data action, A2A or MCP, and why community components were accepted or rejected.

  • 4 · Build

    Configuration-as-code with CX as Code in version control

    Keeps tools, integrations and OAuth configuration reviewable, repeatable and promotable across development, test and production orgs.

  • 5 · Prove

    Negative and failure-path testing

    Attacks agents with prompt injection, out-of-remit requests, failing systems and silent A2A partners before autonomy is granted.

  • 7 · Run & evolve

    Release impact assessments

    Assesses Genesys agentic changes such as model transitions, A2A and MCP availability before tools or autonomy are extended.

See the full QVCCS delivery method

How we deliver

Agentic AI, MCP & custom tooling: your engagement at a glance

  1. 01FrameAgree the tasks AI agents should complete, on whose authority, and which steps must stay with people.
  2. 02DesignTool catalogue, protocols, identity model, audit and guardrails captured in HLD, LLD and ICDs.
  3. 03EngineerDevelopers build data actions, A2A connections and MCP servers in version control under peer review.
  4. 04ProveComponent, SIT, scenario and adversarial testing traced to the RTM before a tightly scoped pilot.
  5. 05Govern & extendSupport aligned to your service model, release impact assessments and step-by-step expansion of tools and autonomy.

The specialists on this work, from our own bench

  • Enterprise Solution Architect
  • Solution Architect
  • Senior Developer
  • Software Developer
  • Systems Integration Tester
  • Senior Platform Practice Lead (Genesys Cloud CX)

Every engagement follows our seven-stage method, with design authority, engineering standards and four-eyes peer review behind it. How we deliver →

Questions

Agentic AI, MCP & custom tooling: common questions

Does Genesys Cloud support the Model Context Protocol (MCP)?

Genesys announced native MCP and A2A interoperability for its Copilots and Virtual Agents in September 2025, and in June 2026 acquired Pinkfish to add MCP-based tool integration, which Genesys says brings 25,000 MCP tools across enterprise systems. A2A connections for agentic virtual agents are documented, while native MCP within Genesys Cloud was expected by early 2027. QVCCS builds bespoke, production-grade MCP servers and designs integrations to adopt native capabilities as they arrive.

How do agentic virtual agents take actions in Genesys Cloud?

Their tools are data actions that call Genesys Cloud APIs, your CRM or other systems in real time, configured with descriptions, pre-instructions, input variables, output validation, masking and error cases. QVCCS specifies each tool in an interface control document with least-privilege credentials and audit logging, so the virtual agent acts reliably and within policy.

What is the Agent2Agent (A2A) protocol in Genesys Cloud?

A2A is an open protocol that lets AI agents on different platforms delegate tasks to each other. In Genesys Cloud, an agentic virtual agent can act as an A2A client through the External A2A Server integration, which is documented as coming soon, delegating work to external agents while keeping ownership of the customer conversation. We design those connections with clear authority and audit.

How do you keep autonomous AI agents safe in a contact centre?

Through layered controls: tightly scoped tools, least-privilege identities, validation inside every tool, Genesys guardrail rules, business rules that require human approval above set thresholds, complete audit logging and active monitoring. We prove these controls with adversarial scenarios traced to the requirements and widen autonomy only as evidence supports it.

Last reviewed

Talk to a specialist about this

We work by introduction. Clients, partners and people introduced to us can reach the right specialist for this topic directly.

Who to contact