← All changes
Breaking changeagentsknowledgedata tablesevaluationsbuild

Agent voice, knowledge retrieval fixes, and Widget SDK 0.7

Agents now answer as your business, knowledge granted by category finally reaches them, rejected record writes are recoverable, and the widget footer changes in SDK 0.7.

What you need to do

Remove `privacyPolicyUrl`, `termsOfServiceUrl` and `disclosureText` from your `AgentKit.init(...)` or `AgentKit.mount(...)` configuration — they are no longer recognised. If your embed snippet pins a version, change `@bymos/[email protected]` to `@bymos/[email protected]`, or copy a fresh snippet from your agent's Embed tab.

New

  • Point of view on an agent's Communication Style card — choose "Speaks as the business" for a brand voice that uses "we" and "our", or "Neutral assistant". Leave it unset and nothing about your agent's wording changes. See answer modes.
  • Send an interim reply, under Activity display — the agent opens with one short line about what it is doing, written for that question in its own voice, while the full answer is prepared.
  • Agents can inspect the structure of tables they have access to — field names and types only, never record contents. Enabled on existing permissions and still editable. See data tables.
  • Run a single test case from its row in a suite, without running the whole suite. It runs once against the current configuration and starts straight away rather than showing the estimate first. It draws on your monthly allowance like any run, but it does not use one of your daily evaluation runs. See evaluations.
  • Test cases can be tagged from the dashboard, suggested from the tags already used in the suite.
  • Cookie preferences in Settings → Privacy & Compliance — one analytics choice, recorded once and honoured across klicforge.ai, this site and the dashboard. See security and privacy.

Improved

  • Agents treat retrieved knowledge as your organisation's own material, so answers say "our rates" rather than "the publicly listed prices". No configuration needed. See core prompt files.
  • A rejected record write now comes back field by field — which field failed, why, and the accepted values for a choice field — so the agent asks one specific question instead of several vague ones. It also stops after three failed attempts and says what it could not do.
  • An agent that describes an action instead of performing it gets one more attempt to carry it out, visible to the contact as it happens, and says plainly if it still could not. See tools.
  • Agents call several tools in one round trip, noticeably faster on answers needing more than one lookup.
  • Knowledge sources created in the dashboard move to combined keyword and semantic search once processing finishes, and a search spanning several sources returns results from each rather than filling up on the highest-scoring one. See knowledge sources.
  • The activity timeline on a conversation shows the gap between steps, so a slow answer can be traced to the step responsible. See conversations.
  • The dashboard sidebar assistant shows its opening line and step-by-step trail while it works, matching the main chat.

Fixed

  • An agent granted a whole knowledge category — a group, or everything at once — retrieved nothing, silently. The same applied to agents addressed by slug, which is what the embedded widget does. Existing grants work from the next conversation; re-enable any source you disabled working around this.
  • A choice field with no options accepted no value, so every write failed. Agents can no longer create one, and now say the field needs options adding. Add them on the table's Fields tab. See field types.
  • Evaluation suites were filed against an agent two different ways depending on how you opened it, so a suite could look as though it had vanished — and the same split broke baseline comparison, froze the wrong configuration into a run, and let a Block publish gate through without running. Existing suites, runs and gates are brought together automatically. See evaluations.
  • The widget's activity trail and its "Worked for Ns" summary now appear; the timer measures the wait until the answer starts arriving rather than counting through the reply.
  • Trail steps name what the agent did — "Searching knowledge base" — instead of an internal tool identifier.
  • A reply quoting two amounts such as "S$250" and "S$500" renders intact; the pair of dollar signs was read as mathematical notation.
  • A signed-in account with no organization gets a clear message instead of the built-in system agents.

Widget SDK 0.7

The composer footer always links KlicForge's own privacy policy and terms of service, alongside a fixed notice that AI can make mistakes. The three options that controlled that area are gone, along with the Footer panel in the widget sandbox — substituting your own links, or hiding the notice with an empty disclosureText, is no longer possible. See the note above for what to change, the widget SDK reference and web widget.