2026-09-02 – 2026-09-08

Skills - a reusable instruction, written once and saved on the agent, that the agent applies on its own when the job calls for it. Skills work in largely the same way as Claude Code Skills.
- Skills live in a new Skills panel on Personalize > Persona, where each one is named, written, switched on or off, edited, and removed.
- The agent picks the skill that fits the question, so the persona no longer has to carry every specialized instruction at once.
- A skill body has the same character limit as the persona, and costs no extra queries - a skill replaces the persona at the right point in the flow rather than adding to it.
- An agent can hold up to 25 saved skills.
- Skills active at one time: 3 on Standard, 5 on Premium, and 10 on Enterprise, where the limit can be raised to 25 on a custom plan.
- Requires Plan & Act. Agents without it are prompted to turn it on. Skills are not available on a Multi-Agent.

The agent can now read a document it has already identified in full, in its original order, instead of searching that document for the best-matching excerpts.
The Actions page now carries a curated CustomGPT.ai connector, so one agent can be pointed at another without pasting an MCP URL. Any agent on the account can be picked, including the same one, and the permissions the connection may use are chosen on the connect screen, independently of what the target agent publishes to its own MCP clients.
Connecting a CustomGPT.ai agent under External APIs no longer asks for an API key. The connection is scoped to all agents or to named ones on the connect screen, and the key behind it is created, moved, and revoked automatically.

Re-verifying a response that has already been verified now returns 400 with a readable message on POST /api/v1/projects/{id}/conversations/{sessionId}/messages/{promptId}/verify, instead of a 500. The in-app route returns the same status code.
agent_capability is now ignored on an agent with Plan & Act enabled, and is reported under ignored in the response rather than counted as updated. Explicit chatbot_model is unaffected.

Fixed the issue where the progress narration in an Enterprise Agents answer ran straight into the answer with no space between them, producing text such as "research on testing.Hello".
Fixed the issue where results returned by an MCP connector led an Enterprise Agents answer to invent details.
Fixed the issue where web search on legacy agents returned out-of-date answers, and where a follow-up question was searched on its own wording instead of the full question.
Fixed the issue where a signed-in user who opened someone else's agent share link was shown the full in-app dashboard wrapped around an agent that is not on their account. That visitor now gets the same standalone view an anonymous visitor gets.
Fixed the issue where a long child agent name in a Multi-Agent chat wrapped onto several lines and broke the chat header layout.
Fixed the issue where searching in the Review Context-Rich Starter Questions modal stopped at 50 results and would not load any more, however many pages matched.
Fixed the issue where signing in through SAML broke when the identity provider rotated its signing certificate. Metadata carrying more than one signing certificate is now imported and used.
Fixed the issue where agent settings could not be saved after a Premium subscription was fully cancelled and replaced with Standard.
Fixed the issue where a Zendesk ticket was skipped entirely when its comments could not be retrieved, even though the ticket's own description was available.