skills and toolsagentsdata tables
Edit your skills with the assistant, and a cleaner agent picker
The assistant now proposes custom-skill edits as change sets you approve, and stops losing track of skill instructions mid-conversation.
New
- Edit a custom skill with the assistant — describe the change ("tighten the wording in the refund skill", "add a required tool") and the assistant proposes a change set you review and approve, the same flow already used for agent edits. Nothing is written to a skill until you approve it, and built-in library skills stay read-only. New skills are still created directly. See writing one with the assistant.
Improved
- Tool call details in evaluation run traces now render arguments and results as syntax-highlighted JSON, with clearer status badges — blocked, error and success are visually distinct — and blocked calls keep their one-click create a fixture shortcut. See reading a run's trace.
Fixed
- The assistant claimed it couldn't see a skill's instructions. It can now always look up the skill library and read a skill's contents, even when a message spans two topics at once — previously a request like "read this skill and update that agent" could route to only one of them and the assistant would say it lacked access.
- The public homepage assistant appeared in workspace pickers. It showed up as a selectable agent when granting Data Table permissions, knowledge access, asset permissions and connections. Platform-internal agents are now excluded from those lists.
- Malformed data-table queries failed opaquely. When the agent built a filter or sort in the wrong shape, the search simply failed. It now gets a specific, correctable error naming the offending filter and what it should look like, so the agent retries correctly instead of giving up or reporting no results.
Changed
- Two platform-maintenance skills, Skill Generator and Skills Management: Admin, are no longer listed in the skill library — they describe how the in-dashboard assistant maintains skills, and that is now the only place they run. They stay attached to any agent that already has them, but their skill-editing instructions no longer apply outside the assistant. If one of your own agents relied on them to edit skills, grant it the Skills tools directly from the agent's Tools tab instead. See the skills catalog.