← All changes
data tablesconversationsagentsorganizationchannels

Reveal encrypted fields, conversation summaries, and table grants that work on their own

Let an operator agent read the real value behind a mask, tell conversations apart from the list, and give an agent a data table without also giving it a skill.

New

  • Reveal encrypted fields — an agent can now return the real name, email or phone behind a mask when someone needs to act on it. It takes two switches: the Reveal Encrypted Field tool on the agent, and Allow reveal of encrypted fields on the table, which is off everywhere until you tick it. Every reveal is written to the table's audit log, the value is shown to the person who asked and cannot be copied into another record or sent anywhere else, and the model itself never receives it. See revealing an encrypted value and AI providers.
  • Summary on the conversations list — a short recap of each ended conversation, so several sessions from the same contact are distinguishable without opening them. It is shown to workspace owners and admins, like the conversation trace. See conversations.
  • Commands on WhatsApp — /start, /reset and /help now work on WhatsApp, the same three a contact can use on Telegram. See commands.

Improved

  • Granting a table is now enough. A read grant lets an agent find and open records, a create grant lets it add them — no data skill required alongside. See permissions.
  • Encrypted fields are masked for agents rather than hidden. An agent can recognise and discuss a record ("Ahmad R.", +65***4567) without holding the real value. See personal data and encryption.
  • View trace in dashboard chat opens the full tool-by-tool record of the conversation you are having with an agent, and those turns now carry model and token figures like the other channels. See trace view.
  • /start and /reset answer immediately with a short greeting from the agent, on every channel that accepts them. See commands.
  • Phone numbers are recognised in international formats as well as local ones, so +65 9123 4567 is removed from a message before the model sees it. See guardrails.
  • Time zones show their GMT offset when scheduling a task, so America/New_York (GMT-4) is pickable without checking. See scheduled tasks.
  • Require a bound contact now appears only under Self mode on a table's access settings, which is the only mode it applies to.

Fixed

  • In a dashboard chat trace, your message appeared below the tool calls it triggered.
  • An agent sometimes said a name or email could not be read when it had just looked the value up.
  • Chat and trace messages showed large blank gaps between paragraphs.
  • The approve button on a change set the assistant proposed sometimes appeared only after a reload.
  • Asking the assistant for a small edit in passing — "make it shorter", "drop that line" — sometimes got a written description of the change instead of a change set to approve.
  • On WhatsApp and Telegram, an agent occasionally replied with a block of raw code instead of doing what was asked.
  • An agent could fail to use an email address or calendar it had just looked up through a connected tool.
  • With Google Calendar connected, an agent reported that a calendar could not be found when the calendar ID was one it had been shown masked.
  • Tool checks in an evaluation scored calls as never made when the agent's progress display was set to minimal or off.