- When an agent answers a person, its
knowledge_search,knowledge_getandknowledge_listcan be narrowed to the knowledge that person can read. That covers a chat reply, a personal agent, a Factory Manager reply and a Workflow Architect builder turn. Workflow steps keep the agent’s own access. See knowledge_search. - The new
knowledge.agentOnBehalfNarrowingplatform setting controls it, and starts inshadow: agents read as before, and each search records how many results the narrowing would have left out. An operator switches it toenforceonce those counts are reviewed. - Under
enforce, an agent that answers a person who can’t be identified, or who has no knowledge access, reads no shared knowledge and says so.
- When the Workflow Architect improves a procedure, the Procedure Builder’s first turn runs for the person whose Architect turn asked for it. It used to be refused because it named the Architect agent, which is no person.
- When the improved version is promoted to provisional, its improvement request now completes too. Before, the request was marked failed (“ended its builder turn without finalizing”) even though the version had been promoted.
- An improvement’s Builder conversation stays open after the version is promoted. An improvement whose sandbox run fails a test the previous version passed is held for review, as the improve regression gate intends.
- The Console’s Knowledge section lists Knowledge, Review, Sources, and an Admin group holding Health, Search quality and Settings.
- What changed is now Review’s Activity tab. Links to the former Learning page open it, with their time window, kinds, scope type and agent filters kept.
- Record collections now open in place in Knowledge’s Collections mode. Links to the former Record collections page open the same collection and record there.
- Any other parameter on those old links is dropped. See Knowledge Base.
- The default
knowledge.signalPrecedenceis now["domain_signal", "operator_rating", "judge", "system"]. A precedence you stored keeps applying. - When an operator’s rating of a run and the run’s domain outcome give opposite signals, the cited artifact goes to the steward queue as a
contesteditem that names both. If the domain outcome was a success, the artifact’s grade doesn’t change. See How outcomes grade a case.
- The Console’s New artifact form has a Needs review box. It’s checked by default for company and global artifacts. A checked artifact goes to the review queue as a Needs review item.
- A steward approves it, which records their verification and keeps the grade, or rejects it, which retires the artifact. Outcomes never close the item.
- Needs review items appear only in the Console. The knowledge API and MCP review tools don’t list or decide them. See Review queue.
- The Knowledge Base no longer switches between Browse and Search by meaning. One box filters the list on your text at once, and after a short pause shows the closest articles and records by meaning above it. The list is headed Text matches while you search. See Search and browse.
- The search text is in the URL as
search. Older search-by-meaning links (mode=meaning&q=…) open the same box with their text. Collections is unchanged.
- An artifact created with Needs review is hidden from agents and API clients until a steward approves it. That covers
knowledge_search, auto-inject,knowledge_get, the knowledge API’s search, ask, get, cases, export and learning feed, and the MCP knowledge tools. - People who can read its scope still see it in Knowledge, with a Pending review badge. Approving it makes it visible to agents. Rejecting it retires it.
- The review queue now names who decided an item. See Review queue.
- When a personal assistant’s memory earns automatic promotion, the new artifact goes to its owner’s personal scope, even if the memory was written in a project or during a project run. If the assistant’s owner has no worker, the memory isn’t promoted.
- Promoting a memory by hand from the Console works as before. See How a memory becomes a case.
- An agent’s
knowledge_recordcall can setkindtolesson,policy,how_to,reference,faqornote. It’s stored as the artifact’s kind, the same field the Console’s New artifact form sets, so you can filter by it in Knowledge. - Kind doesn’t change search or retrieval. A rewrite that leaves out
kindkeeps the artifact’s current kind. Seeknowledge_record.
- When an operator stops people, API key or agent access, every refusal now says so in a fixed sentence without the operator’s reason, for example “An operator has stopped people access to Forge.” Stop everything reads “The platform global stop is engaged.” This applies to the Console’s people-door error, Slack and email replies, the Platform API and inference
503governance_door_stoppedanswers, recorded decisions, step and run errors, Builder and Architect messages, and tool results. - The reason you type when you stop a door or Forge stays on Emergency stops, for operators with the switches permission. A reason can name another customer or an incident, so it is not shown to refused callers or to organization readers.
- This replaces the 2026-10-13 change that gave the operator’s reason on every surface. Decisions recorded before this change keep the text they were recorded with.
- Roles draws Base, Builder and Operator as an indented ladder that says what each step adds in words, lists the other roles alongside and the agents’ roles apart, and compares Builder with Operator by default, with the added permissions as chips. With See access, each role shows how many holders it has at your company; without it, no count is shown.
- Holders groups holders in bands on a rail by where each grant was given, as cards that say whether the role was given here or inherited, and can be searched. A holder’s page opens on how each grant reaches the scope, then why the access exists, then what it allows.
- Access check asks with actor and permission choices and questions phrased for the actor, answers with a verdict card, draws where Forge looked on a rail, and keeps the company inventory closed until asked for.
- Delegation draws one lane per grant (workflow version, connection, company) with its status (an expired grant reads Expired), grantor, expiry, Review and Revoke…, and says that a stored grant is not live credential access.
- History has one toolbar (Find by Reference, Outcome, Group by, Window, More filters) and one Export menu. Less common filters sit under More filters; a filter set there or by a link shows as a named chip you can remove.
- A decision opens in a drawer beside the list: its outcome, its checks as one track, what to do next, and the record as one grid of recorded facts, with the facts the record didn’t keep named once.
- Reports offer the last 7 or 30 days or a custom window, and answer with a bar per row and its change against the window before; the table is one click away. An unavailable report explains why it shows no count and points to History.
- The gate explorer draws a red bar beside each failure count, and Show failures in this view opens History at that check’s recorded failures only. On a phone each check is a card.
- On the overview, a procedure decision names its procedure, coverage notes on Needs attention read in plain words, a page with no organization name reads “Company work” and “Recent decisions” rather than a placeholder, and the stage legend’s example matches the first stage.
- The overview opens with a plain sentence about what needs a person, then up to three counts from what loaded: items waiting for a person, agents blocked by an accountability finding, and the global stop. A count from a partial read says “at least” (“At least three decisions need a person.”).
- The path panel draws its stages in one row in the gate registry’s order, with the global stop and Agents door state as badges. The separate control posture card is gone; Emergency stops is linked from the panel.
- Needs attention sits beside where work was refused, which now shows recorded failures per check as bars. Agent accountability lists up to five agents with a finding, most severe first, counts the rest with a link to the accountability register, and says how many of the loaded agents are in good standing.
- Needs attention rows each have their action (Review request, Explain hold or Review agent). Recent decisions name who acted, with agents named as in History, and the check that decided each one, with its path and check number when the check runs on one path only. Times are local: the time alone for today, the date and time otherwise.

