Wonderful

Agents connect to the systems the work runs on, reading data, triggering actions and updating records directly in CRMs, ERPs, policy systems and data platforms. Engineers build on the underlying integrations, APIs and tools, and enterprise MCP gateways control what each agent is allowed to call.

Sourcewonderful.ai/blog-articles/wonderful-an-enterprise-platform-to-turn-ai-ambition-into-agents-in-productionread 2026-10-03

Each agent is a composition of skills and loads only the ones the current task needs, so a single agent can carry a multi step process across domains and make decisions along the way. An orchestrator and managed workflows run complete business processes from the moment work arrives until it is resolved, and work can be handed between agents on different platforms over the A2A protocol.

Sourcewonderful.ai/agentsread 2026-10-03

What the organization already knows grounds every agent: its data, documents and policies become the shared context agents work from, read under one set of permissions. Knowledge bases sit in the chain each agent draws on, and each skill can carry its own knowledge for a specialized task. How documents are indexed and refreshed is not described publicly, so the maintained store is an inference from these platform descriptions.

Sourcewonderful.ai/agentsread 2026-10-03

A policy layer monitors every conversation for attempts to reach sensitive data or disclose what an agent should not, and teams configure the response, such as transferring to a human or ending the conversation. Agents pause where human review is required and resume once the work is cleared, and Wonderful Apps puts approve and decline controls in front of the people supervising each agent.

Sourcewonderful.ai/blog-articles/the-hard-part-of-ai-agents-isn%e2%80%99t-building-themread 2026-10-03

Wonderful states SOC 2 Type II and ISO 27001:2022 certification, PCI DSS and GDPR compliance and EU AI Act alignment, verified independently each year, plus third party penetration tests twice a year. Permissions decide who can build, edit and ship agents down to the individual resource, and the AI Gateway sets which roles or teams may use which models. The detailed control list sits in a Vanta trust center at trust.wonderful.ai.

Sourcewonderful.ai/ai-osread 2026-10-03

Conversations, decisions and tool usage are observable by default, so a team can see why an agent behaved as it did and not only what it did. Failures can be traced to the exact moment in a conversation and tagged, alerts flag tool failures and unusual volumes, and an immutable audit trail covers the environment, including every action an agent takes inside a legacy system.

Sourcewonderful.ai/blog-articles/the-hard-part-of-ai-agents-isn%e2%80%99t-building-themread 2026-10-03

State carries through a run: an agent that pauses for human review picks up where it left off once the work is cleared. The AI OS page says every decision builds a record the platform learns from, but no memory layer that agents read across separate conversations is documented, and no scope or retention period is given.

Sourcewonderful.ai/blog-articles/computer-useread 2026-10-03

Four deployment models are offered: multi tenant SaaS with logical isolation, dedicated single tenant infrastructure, bring your own cloud on AWS, Azure or Google Cloud with data kept inside the customer's organization, and a fully air gapped on premises install maintained on site by Wonderful's engineers. Data residency options are published in the trust center.

Sourcewonderful.ai/deploymentread 2026-10-03

A catalog lets a customer build an agent, skill or tool once and reuse it across the organization, and reusable skills package a procedure for any agent that needs it. Those are assets the customer's own teams assemble. Industry pages describe workflows such as KYC and credit servicing, but they are taken into production by deployment teams rather than offered as prebuilt agents or templates a customer adopts.

Sourcewonderful.ai/agentsread 2026-10-03

Work reaches the agents through voice, chat, email, documents and embedded experiences, with the same skills applying on every channel. Live deployments include inbound calls at a national telecom, appointment booking at a bank and collections across 2.5 million clients, and an agent waiting on a human approval resumes when the approval arrives. Schedule and webhook triggers are not described on public pages.

Sourcewonderful.ai/agentsread 2026-10-03

Nothing built is locked to one model: the platform routes each task to the model that fits, and new models drop in without rework. Admins keep control through the AI Gateway, which sets which roles or teams can reach which models, caps budgets and limits premium model use, and the company's CTO says customers choose the best models for each workload. A per agent model picker is not shown on a public page.

Sourcewonderful.ai/ai-gatewayread 2026-10-03

The whole platform is exposed through an API documented in a published Swagger file with hundreds of endpoints, and it can run headless, so customers can build their own interfaces or use it purely as an orchestration layer. Every agent, skill, tool, governance setting and app can be exported through that API, and a CLI is part of the AI OS. The API reference sits behind a customer login at app.wonderful.ai/docs.

Sourcewonderful.ai/blog-articles/open-by-default-competitive-by-designread 2026-10-03

Changes are tested before they ship: generative evaluations simulate real interactions and score whether the agent met the intent behind each request, catching regressions and edge cases. In production, automated checks surface drift and failures, and A/B tests split traffic between two agent versions so the better one can be promoted and the other rolled back.

Sourcewonderful.ai/blog-articles/fast-iteration-as-a-strategyread 2026-10-03

Legacy systems with no API are reachable: agents log into software such as Oracle, Epic or SAP, move between screens, fill in fields and submit forms the way staff do. Each one runs in a managed virtual machine with its own credentials and a controlled session, every action is logged, and a single view shows which sessions are active or waiting on human review.

Sourcewonderful.ai/blog-articles/computer-useread 2026-10-03