# Reviewing what your team teaches

When someone on your team corrects an agent, it arrives with you. This guide
covers what to do with it, how to undo something that went live, and how to
decide how much your agents may publish without asking.

> **Reading this before you have a workspace?** Everything below describes
> what a reviewer does inside Karmaflow.ai. The permissions named are the ones a
> workspace admin grants — you do not need them to read on.

## What you'll need

- **`quality:events:read`** to see the quality stream and the Curator Inbox
  (the seeded Manager, Member and Viewer roles all carry it).
- **`quality:events:manage`** to promote, dismiss, or change the sharing
  setting.
- **`knowledge:brain:read`** to open concepts, **`knowledge:brain:manage`** to
  publish or roll back, **`knowledge:brain:govern`** to change autonomy
  settings.

If a page below is missing from your menus, it is a permission — not a missing
feature.

---

## Where corrections arrive

There are three doors onto the same work. Use whichever suits how you are
working.

| Door | Path | Best for |
|---|---|---|
| **The conversation itself** | Open the call/chat/email | Handling one thing you were told about |
| **Training threads** | Agents → Workflows → Quality Events → Training threads | Seeing everything your team has taught, with each thread's receipt |
| **Quality Events** | Agents → Workflows → Quality Events | Working the whole stream, and the Curator Inbox |
| **Tenant Brain** | Agents → Workflows → Tenant Brain → Approvals | Reviewing drafted knowledge waiting to publish |

## Handling a correction in the conversation

Open the conversation and find the thread. You get the whole exchange: what
they quoted, what they said, and what the AI already did about it — including
who left it.

You can reply in plain language to direct the fix, exactly as the original
author could. The AI shows its work as it goes:

![Three completed steps listed in order — Searching the brain, Reading knowledge, Drafting a correction — each with a green check.](/help/assets/knowledge/train-agent-working-steps.png)

Then it reports the outcome. If your workspace's autonomy settings allowed it
to publish, it says so plainly and names the authority it acted under:

![A green card reading "Live — published under earned autonomy" above a summary of the change that was made.](/help/assets/knowledge/review-live-card.png)

The before-and-after on the original thread:

![An amber notice reading "Draft created from this — waiting on an admin to review it" with a link to open the draft.](/help/assets/knowledge/review-before.png)

![A green notice reading "This became knowledge — policies/service-levels.md. Thank you for teaching it."](/help/assets/knowledge/review-after.png)

The person who left the correction sees that same receipt on their own **Train
the AI** page — you do not need to tell them it was handled.

## The Curator Inbox

**Agents → Workflows → Quality Events → Curator Inbox**

The Curator is the AI that works feedback about **how agents behave** — the
prompt-change plane. It reads the feedback in context and proposes what should
change about an agent's instructions.

It runs as a series of steps you can follow:

![The Curator's analysis stepper, showing its progress through analysing a piece of feedback.](/help/assets/knowledge/review-curator-stepper.png)

And it writes up what it found:

![The Curator's written analysis of a piece of feedback.](/help/assets/knowledge/review-curator-analysis.png)

> **The Curator is allowed to disagree with the feedback, and that is a
> feature.** In our own testing it declined to act on a style correction,
> saying it was not well grounded in the transcript and that applying it could
> "overfit to a detail not evidenced". It proposed no changes. An AI that
> refuses a weakly supported instruction is protecting your agents from being
> tuned by a single bad day — read its reasoning before overriding it.

Entries must be **analysed before they appear in the inbox** — the queue lists
feedback the Curator has already worked, not raw submissions.

## Publishing, and undoing

Corrections that become knowledge are ordinary revisions of a concept, so they
are reviewed, published and undone through the knowledge base.

**Agents → Workflows → Tenant Brain →** open the concept **→ History**.

### Reading the history honestly

Each revision shows **two separate times**, and the difference matters:

- **Written** — when the revision was created.
- **Published** — when it actually went live and agents began reading it.

They are often hours apart, and a revision can be written and never published
at all. Badges make the state explicit: **live now**, **was live**, **never
published**, or **no publish on record** for revisions that predate publish
tracking.

> **Never read a "Written" time as a publication.** A revision existing does
> not mean it ever reached your agents.

Use **Compare** on any revision to see exactly what changed against the one
before it, before you act on it.

### Two ways to undo — they are not the same

| | **Revert** (on an audit row) | **Restore** (in History) |
|---|---|---|
| Where | Tenant Brain → Approvals, and inline in History while the window is open | Version History, any revision |
| When available | Only while an auto-publish's review window is open | Always |
| Goes back | Exactly one step | Any revision you pick |
| Effect | Immediate — live in one click | **Stages** an edit; needs a separate **Publish** |

**Restore does not revert.** It loads the old content into the editor as an
unpublished edit — your agents keep reading the current version until you
publish. The banner says so explicitly:

![An amber banner reading "v28 is loaded into the editor as an unpublished edit — this is not live yet. Your agents are still reading v30. Publish to complete the rollback." with a "Discard it instead" link.](/help/assets/knowledge/review-restore-banner.png)

If you change your mind, **Discard it instead** puts everything back.

When you do publish a rollback, a confirmation names the live claim you are
about to contradict and notes that publishing anyway is recorded against your
name. Read it — it is the last chance to catch a wrong target.

**Nothing is ever erased.** A rollback is recorded as a new revision, with the
change and the undo both visible in history, and the undone revision keeps its
own Restore control so it can be brought back.

## How much should agents publish on their own?

**Agents → Workflows → Tenant Brain → Governance**

This is where you decide how much autonomy your agents have. Options run from
reviewing everything, through publishing only what a proposer has earned trust
to publish, up to publishing verified high-confidence knowledge automatically
with a review window for undo.

Two things worth knowing before you change it:

- **It changes what your colleagues see.** Under higher autonomy, a
  manager's correction may go live during the conversation; the thread reports
  that honestly rather than promising a review that will not happen.
- **Not every automatic publish is individually queued for you.** A sampled
  portion gets a keep-or-revert window; the rest publish cleanly. That is the
  point of autonomy — but it means the queue is not a complete record.

To see everything that published automatically — sampled or not — use the
**Auto-published** list on the Approvals tab. It covers the last 30 days,
needs no action, and offers the one-click Revert where a window is still open.

## Sharing anonymised quality patterns

**Agents → Workflows → Quality Events**

Near the top is a checkbox: **Share anonymized quality patterns with the
platform**.

![The "Share anonymized quality patterns with the platform" checkbox with its explanatory text.](/help/assets/knowledge/review-sharing-setting.png)

It shares **patterns**, not content — the shapes of what tends to go wrong,
used to improve the platform for everyone. Read the sentence beside it for
exactly what leaves your workspace and what never does. Changing it requires
`quality:events:manage`; everyone else sees it read-only with a note saying
who can change it.

## Troubleshooting

**Someone says they left a correction and I can't find it.**
Open **Quality Events → Training threads** — every thread anyone has opened is
listed there with its receipt, and each row links back to its conversation.
(The Curator Inbox is not the place to look: it only lists feedback that has
already been analysed.) If it is still missing, check whether the thread is on
an agent type your role can read — the listing only shows agent types you hold
read access on.

**A correction published when I expected to review it.**
Check **Governance**. Under higher autonomy, a correction from someone with
publish rights can go live immediately. The Auto-published list on Approvals
shows what did.

**I reverted something but the person's receipt still says "became knowledge".**
That receipt describes what happened at the time and does not currently update
when a concept is later rolled back. If the correction mattered to them, tell
them directly.

**I rolled back but agents are still using the old text.**
You staged a Restore without publishing it. Look for the amber banner — the
rollback completes on **Publish**.

## Next steps

- **[Quality Events and the Curator queue](/help/knowledge/training-curator)** —
  the two surfaces in depth: the full event stream with its filters and
  promote/dismiss controls, topic muting, and the per-agent **Feedback Review**
  workflow where a prompt change is approved and applied.

## Related

- [Agents that learn from your team](/help/knowledge/training-overview) — how the whole loop fits together.
- [Training your AI agents](/help/knowledge/training) — the overview and role model.
- [What happened to your feedback](/help/knowledge/training-outcomes) — the receipts your team sees.
- [Brain Browser walkthrough](/help/knowledge/overview) — operating the knowledge base itself.