Context
Context (in the sidebar under Knowledge Base) is where you see what your AI agents actually know — and whether that knowledge helps.
As agents implement tasks, they capture learnings: reusable patterns, decisions, and pitfalls written down so the next task doesn't relearn them. Context turns that growing pile into something you can inspect and judge: every learning carries its track record — how often agents applied it, how often it helped, and how often it misled.
The lenses
One knowledge base, several ways to cut it:
| Lens | Question it answers |
|---|---|
| Most used | What do agents lean on the most? |
| Recently used / Recently added | What's the live pulse — and what just entered? |
| Needs review | What's been applied a lot but rarely helps? (With a count badge, so the backlog is visible at a glance.) |
| By team / By domain | Which team or knowledge area owns how much? |
Search covers titles and paths; the header shows workspace totals (learnings, applications, overall hit rate, needing review); and every view — lens, filter, opened learning — is a shareable link that survives a refresh.
A dossier per learning
Expand any learning in place to see the full picture: helped / no-effect / misled tallies as a traffic light, the owning team, the task it was captured in, when and where it was last applied — plus an Applied in feed listing each task it was used in, by whom, with what outcome and note, and the learning's full text.
Where the numbers come from: whenever an agent applies a learning during a task, it logs the outcome — helped, no effect, or misled — as part of the Task Implement workflow. The signal builds on its own; nobody fills in a form.
Needs review — and what to do about it
The platform raises the flag, not a heuristic you have to run: a single misled outcome flags a learning immediately; otherwise it takes a meaningful sample (several applications with something other than "helped" among them). Each flagged card has a ⋯ menu:
- Dismiss — clears the flag; only evidence after your review can re-raise it, so the same old history doesn't instantly bring it back.
- Mute — turns flagging off for that learning for good.
- Delete — removes the learning so agents stop retrieving it (its application history is kept). Works on any learning, flagged or not.
Every action asks for confirmation first.
Why this matters
Knowledge bases usually grow by addition only — nothing ever proves its worth, and nothing gets retired. Context closes the loop: what helps rises to the top, what misleads gets flagged, and you curate with evidence instead of memory. The docs and instructions your agents read stay in Docs and AI Workflows; Context is where their accumulated experience lives.