Skip to main content

Creatio AI Studio updates: September 16, 2026

In this update, agents that AI Twin builds tell people when a tool fails. The agent editor explains autosave errors and starts version history at version 1. Agents without their own run time budget can run for up to 600 seconds. Webpage knowledge sources keep serving content while they refresh, and the execution timeline shows each step in order. Voice agents also announce a transfer before it happens. Learn more about all the new capabilities below.

AI Twin​

Tool failures reported by agents that AI Twin builds. When AI Twin creates an agent or edits its prompt, the prompt tells the agent how to report failures. When a tool call fails or data cannot be read, the agent says in plain language what failed and what it means. This way, people who use the agent know when an answer is incomplete.

Agent design and configuration​

Version history starts at version 1. Some agents are deployed before they are first published, for example, a personal agent that AI Twin creates and deploys for you. Their version history starts at version 1. The deployed version is listed under Versions, and later version numbers match the versions you deploy. Agents whose history already starts at version 2 keep their numbering.

Autosave errors in the agent editor. When the agent editor cannot save your draft, it shows the reason in a message. For example, the selected model may not be able to run one of the agent's built-in tools. The tooltip of the Publish button repeats the reason, starting with "Fix the autosave error before publishing." The tooltip clears after the next successful save.

Draft reset when there is something to reset. The Reset draft to current release action appears only when your draft differs from the current release. If you reset a draft while the editor is still saving an edit, the editor shows the content of the current release after the reset.

Improved text contrast. Read-only fields, for example, on a published agent or skill, use higher-contrast text. This makes a published configuration easier to read.

Right-to-left text in the agent preview. The Text preview of an agent shows mixed right-to-left and left-to-right text in reading order. An example is Hebrew with English terms, links, code, or tables. This lets you test agents that answer in right-to-left languages.

Longer default run time budget. An agent without its own run time budget can run for up to 600 seconds. This gives long tasks more time to finish. The Run time budget (seconds) field in the Runtime limits block of the Advanced tab shows 600 until you set a value. Agents with their own run time budget keep it.

Models​

Built-in tools while the model catalog loads. Sometimes the built-in tool support of the selected model is not known yet. Until it is known, you can turn on only Web search in the Built-in model tools group of the Tools tab. Other built-in tools show Waiting for the model catalog, and a tool that is already on can still be turned off. A new agent starts with Image generation only when its model supports it.

Skills​

Creatio data skills aligned with the Creatio tools. The out-of-the-box skills for Creatio data follow the current behavior of the Creatio tools. These skills query, create, update, and analyze records, explore the data structure, and trace stage transitions. Agents that use them can sort by any column and follow a relationship from either object. They can also page through aggregated results and check a new record for missing required fields from its first validation. The date filtering and reporting periods skills take the current date from the platform. Copies of these skills that you edited keep their content.

Knowledge​

Webpage source refresh keeps current content. When you re-fetch a webpage knowledge source, every page is fetched again. A page whose text has not changed is not indexed again. The Web Content Fetch Sessions table on the Web fetch tab counts those pages under Skipped. Each page is replaced only after its new version is fetched, so a page that cannot be fetched keeps serving its current content. On the URLs tab, Last extracted shows when the served version of a page was extracted. Last attempt shows when the page was last fetched.

Fair share of indexing capacity. Knowledge source indexing is scheduled fairly. Part of the indexing capacity is kept for a refresh or re-index that you start yourself. This way, large indexing jobs do not hold up your work. When no capacity is free, indexing waits and starts on its own.

Search through long attached files. When an agent searches a long attached file, it continues from where the previous page of matches ended. This way, it can find matches in any part of the document.

Channels and integrations​

Configurable handoff announcement. Before a voice agent transfers a call to a handoff target, it tells the caller that it is connecting them. The transfer starts after the sentence ends. You can set the sentence for each target in the Announcement field on the Handoff targets page of the Telephony section. For example, enter "One moment, I am putting you through to our billing team." The agent says the text as written, without translating it. If Announcement is empty, the agent words the sentence itself in the language of the call. An announcement can have up to 500 characters.

Context for outbound calls. An outbound call can be requested with call context, for example, the case the call is about. The voice agent gets that context as soon as the person answers. It can then refer to it from the first turn of the conversation.

Clear consent prompts. An agent can work in Creatio on your behalf with the access you granted. When Creatio stops accepting that access, the agent asks you to authorize again and gives you a link. The link opens the Creatio environment that the conversation runs in. The agent then uses the access you grant there for its next request.

MCP tool errors in run history. When an MCP tool reports an error, the call is recorded as failed. The run's Execution timeline and its tool error count show the outcome that the tool reported. An MCP server can send a tool description longer than 2,000 characters. The tool is still added, with its description shortened to the limit and marked Description shortened, so the agent can use it. Descriptions that you enter yourself must stay within the limit.

Runtime and deployment​

Rate limits per agent run. The platform limit on model requests applies to each agent run separately, within an overall limit for your organization. This way, a busy run does not use up the requests that other runs need.

Observability​

Clearer execution timeline. The Execution timeline of a run on the Agent Observability page lists steps in the order they happened. Each step's bar starts when the step started, including a model call that failed after a long wait. When the agent works through multiple cycles in one run, a divider marks each cycle, for example, "Iteration 2." The divider shows the step count and duration of the cycle. Each PII protection check appears next to the model call that it came before. In the list of runs, a run with failed tool calls shows the number of tool errors and counts toward Needs attention. The Tool errors filter finds runs With tool errors or Without tool errors, including runs that completed.

Usage and budgets​

Usage totals for every range. The Usage analytics page shows your organization's totals for every range, including Last 30d. When a section cannot load its data, it says "Couldn't load usage data" and offers Retry.