Skip to main content

Creatio AI Studio updates: September 25, 2026

In this update, you can hand an AI Twin conversation to another agent and keep everything said so far. While an agent works, a short status line shows what it is doing, and each error notice starts with a clear title. Agents can search inside attached PDF, Word, Excel, and PowerPoint files. You can limit where a telephony integration may call, and voice agents transfer callers to the target they ask for. Reading run step content is now a separate permission, which only administrators have until you add it to other roles. Learn more about all the new capabilities below.

AI Twin​

Conversation context kept when you switch agents. You can hand an AI Twin conversation to another agent and keep its context. When you select another agent in a conversation, that agent sees the earlier messages, including the results of tools used before the switch. If you switch back, the previous agent continues the same conversation. This also works in conversations started before this update.

Generated conversation titles. AI Twin gives each new conversation a short title based on your first message. This helps you find a conversation in the list by its topic. Until the title is ready, the list shows the beginning of your message. Administrators can choose the model that writes titles in Conversation titles on the Platform LLM page of Settings. If no model is selected, Use Luna (default) applies.

Error notices with their own titles. When AI Twin cannot complete a request, the notice starts with a short title that says what happened. Examples are "Agent unavailable," "Session expired," "AI usage limit reached," and "Model service unavailable." The explanation follows the title. A request stopped for a known reason, such as a usage limit, a missing permission, or a pending approval, shows as a warning. A service failure shows as an error.

Short status line while an agent works. While an agent works on your request in AI Twin, a short line below the answer says what is happening. For example, it reads "Looking up data…," "Creating a record…," or "Using an integration…." Otherwise, it reads "Thinking…." While the answer appears, it reads "Typing…." A request that waits for your decision shows "Awaiting approval." A request you stop shows "Canceling…" and then "Canceled." The line disappears when the answer is finished or the request fails.

Agent list follows deletions. When an agent is deleted in Creatio AI Studio, it disappears from View all agents in AI Twin without a page reload. This applies to personal and enterprise agents, pinned or not. Your conversations with the agent stay available.

Agent design and configuration​

Completed steps in the agent preview. When an agent finishes a reply in the Text preview of the agent editor, every step of the reply is shown as completed. This includes steps that use Web search or Code execution, so you can see when the agent is done.

Skills​

Delete individual files from a skill. On the Skill tab of a skill draft, each file you can delete has its own delete control. You confirm the deletion before the file is removed from the draft. You cannot delete the SKILL.md file or folders. If another change to the skill was saved first, the editor asks you to reload the skill and try again. Files of vendor skills and of published versions cannot be deleted.

Steadier answers from the Creatio data skills. The out-of-the-box skills for Creatio records, folders, and stage history give more consistent answers. When a question is about a team, the agent uses every owner that your organization structure returns. Some questions about earlier periods use a folder that excludes records, for example, closed-lost deals or a country. The agent then decides from your question whether each exclusion applies. If your question does not make it clear, the agent asks you or gives both answers. It also says which exclusions it left out. When a value you enter for a new record does not meet the filter of a lookup, the agent explains what the filter checks. It then asks you for a suitable value.

Knowledge​

Search inside attached documents. An agent can find a phrase in an attached PDF, Word, Excel, or PowerPoint file and quote the matching passages. It works the same way as for text files. This way, the agent can answer a question about one value without reading the whole document. When the search service is busy, the agent says so and can try again. The agent also says when a document is too large to search in time.

Agent evaluation​

Delete eval cases and keep run history. You can delete any eval case, even one that has already run, and keep the results of earlier runs. On the Cases tab of the agent's Evals tab, click the delete icon in the row of the case. The case is removed from the agent's cases and from all its suites. The runs it was part of keep its results on the Runs tab.

Channels and integrations​

Outbound destination policy for telephony. You can see and limit where your telephony integration may call. Use Destination policy on the integration's tile on the Integrations page of the Telephony section. In Allowed destinations, select Domestic, International, or both. In Forbidden destination prefixes, add dial prefixes that the integration must never call, such as +1900. Prefixes marked Built-in or Installation are inherited and always apply. The built-in prefixes also cover national premium-rate numbers in Russia, the United Kingdom, Poland, Indonesia, Ukraine, and the United Arab Emirates. The policy works together with the permissions of your carrier. It can only narrow where the integration may call, never widen it. Each save appears on the Audit Log page as Outbound destination policy changed. If your organization has multiple telephony integrations, the policy is enforced from one of them, and the Destination policy dialog of the others says so.

Transfers to the handoff target the caller asks for. A voice agent connects the caller to the handoff target they ask for, for example, "put me through to sales." It matches the request with the Name of the active targets on the Handoff targets page of the Telephony section. The caller hears the announcement of that target before the transfer. If the caller does not name a target, the agent uses the only active target. If there are multiple active targets, it uses one whose Scope is All voice channels. Targets are read when the call starts, so changes apply from the next call.

Announcement when a voice agent cannot join a call. If the voice agent does not join an answered inbound call within 10 seconds, the caller hears that the call cannot be connected. Then the call ends. The call appears in the Calls list of the Telephony section with the Completed status and its reason. The message is in English.

Voice channel warning on Web Chat. Sometimes the voice channel that handles the calls of an agent is switched off. The Capabilities tab of the Web Chat channel then shows a warning with the name of that channel. Click Open channel to switch it back on.

Creatio access kept through connection interruptions. An agent that works in Creatio on your behalf renews its access from time to time. If Creatio is briefly unreachable at that moment, the agent keeps the access you granted. It does not ask you to authorize again.

Runtime and deployment​

Runs wait for long model answers. Sometimes the model sends nothing for 30 seconds or longer while it prepares a large answer. The agent run keeps waiting and completes when the answer arrives, as long as the run time budget allows.

Recoverable rate limiting. When an agent run reaches the platform limit on model requests, it waits until the limit resets and then continues. This works as long as enough of the run time budget is left.

Dependent environments named when an environment cannot be deleted. You cannot delete an environment that another environment uses as its promotion source. When you try, the message on the Environments page lists those environments. This shows you which promotion source to change or remove first.

Observability​

Run step content is a separate permission. You can let more people review run history without showing them the data inside it. To read step content on the Agent Observability page for a run you did not start, you need the Read run step content permission. Step content includes tool input and output, the conversation with the model, policy evidence, and run error text. Without the permission, you still see every step, count, duration, and status, and the content reads "Payload hidden for your current role." Out of the box, only the Administrator role has this permission. Users with other roles, such as Developer or Individual Contributor, see step content only for runs they started. To change this, add the permission to their role on the Roles tab of the Security section.

More complete run timeline. The Execution timeline of a run on the Agent Observability page records more steps, with the input and output of each. It records each knowledge source search. It records each use of the built-in tools that list, read, or search attached files. It also records each time the agent opens a skill it opened before. A failed call of an attached-file tool counts as a tool error. In the details of a model call, Available skills lists every skill offered to the model. Runs recorded before this update do not change.

Interrupted model calls shown as failed. A model call can be cut off, time out, be cancelled, or end with a provider error. The Execution timeline of the run then shows it as a failed step with its error, so you can see which call stopped the run.