Agents that learn from your team

Most AI products give you a thumbs-up button. The rating goes somewhere you cannot see, changes something you cannot name, and you find out whether it worked by waiting to see if the mistake happens again.

KarmaFlow trains differently. Someone on your team highlights the exact words an agent got wrong, types what is actually true, and gets back a receipt naming what changed — a concept in your knowledge base, a prompt change filed against an agent, or an honest "that was specific to one conversation, nothing was saved." The correction is versioned, attributed, and revertible.

What It Covers

Correction where the mistake happened. Any agent-authored text is trainable in place — voice call transcripts, chat sessions, SMS threads, sent email (subject and body as separate targets), task-agent output, and agent activity surfaced on CRM records. The gesture is identical on all of them: drag-select, click the chip, say what is true.

Routing that admits three answers. An AI reads each correction and decides where it belongs — a durable business fact, a change to how an agent speaks, or something true only of that one conversation. It is allowed to conclude "nothing should change here", and it says so rather than quietly writing something.

A knowledge base you can read and undo. Corrections that become facts land in the Tenant Brain, your workspace's knowledge in the Open Knowledge Format (OKF). Every concept carries citations, a revision history that separates written from published, side-by-side Compare, and a Restore that stages an old revision as an unpublished edit rather than silently swapping what your agents read.

Autonomy you set, and can audit. Governance settings decide how much publishes without a human — from reviewing everything, through publishing only what a proposer has earned trust to publish, up to auto-publishing verified high-confidence knowledge with a window in which one click reverts it. Every surface reports which of those actually happened instead of promising a review your settings will not perform.

A second plane for behaviour. Feedback about how an agent speaks goes to the Curator — an AI that reads the feedback against the transcript and the agent's current prompt and proposes bounded edits. It is allowed to disagree. In our own testing it declined a style correction as insufficiently grounded in the transcript and proposed no changes.

A quality stream underneath all of it. Message feedback, whole-interaction feedback, CSAT, session audits, performance scores, ticket insights and not-durable verdicts land in one stream with sentiment and per-agent attribution, feeding both the Curator and the knowledge distiller.

flowchart TD
    subgraph S["Where agents work"]
        V["Voice calls"]
        C["Chat + SMS"]
        E["Sent email"]
        T["Task runs"]
        R["CRM activity"]
    end
    S --> H["Someone highlights the exact words<br/>and types what's actually true"]
    H --> J{"What kind of<br/>correction is it?"}
    J -->|"A durable business fact"| K["Tenant Brain — cited,<br/>versioned, revertible"]
    J -->|"How the agent speaks"| P["The Curator proposes<br/>a bounded prompt edit"]
    J -->|"True of one conversation"| N["Noted and closed —<br/>nothing published"]
    K --> G["Governance decides:<br/>publish now, or queue for review"]
    P --> A["A reviewer ticks the edits<br/>they accept, then applies"]
    G --> Q["Every agent reads the<br/>corrected knowledge"]
    A --> Q
    K --> RC["A receipt on the original<br/>message, naming what changed"]
    P --> RC
    N --> RC
    Q --> S

Why Teams Use It

The people who spot mistakes are the people who can fix them. Training needs no permission at all. The rep reading the transcript does not file a ticket for an admin — they correct it where they found it. Reviewing and publishing stay permissioned; noticing does not.

Nobody has to learn the knowledge base. Correcting an agent does not require knowing which concept holds the fact, or what a concept is. The AI finds where the fix belongs.

Corrections stop being a matter of faith. Every one ends in a named outcome. "Became knowledge — policies/service-levels.md" is checkable; a thumbs-up disappearing into a dataset is not.

One mistake is fixed once. A fact corrected on a voice call is read by the chat agent, the email strategy, and the task agent from that point on. The loop compounds across channels rather than per-agent.

Members contribute without being able to break anything. A member's correction always lands as a draft for review. That is what makes it safe to open training to the whole workspace.

Business Outcomes

What changes Why it follows
Wrong answers stop repeating across channels The correction lands in shared knowledge every agent reads, not in one agent's prompt
Corrections come from the front line, not a backlog No permission gate on training; no ticket to file
A bad correction is recoverable Versioned revisions, Compare, and a Restore that stages rather than swaps
Style drift is handled separately from facts Behaviour feedback routes to the Curator and the agent prompt, never into your knowledge base
Autonomy can increase without losing the audit Auto-published changes are listed for 30 days with one-click revert while the window is open
Your knowledge base stays clean One-off details are judged not durable and deliberately not saved

What's Different From a Thumbs-Up Dataset

Rating-only feedback KarmaFlow training
What you point at A whole response The exact words, quoted back to you
What you say Good / bad What is actually true, in plain language
What you get back Nothing A receipt naming the concept, the prompt change, or the decision not to change anything
Where it goes An opaque dataset A named, cited concept in your own knowledge base
Can you undo it No Version history, Compare, Restore, and revert-in-window
Who may do it Usually admins Anyone in the workspace; publishing stays permissioned
Portability None Open Knowledge Format — the knowledge is yours to export

Ratings still exist and are still useful — a bare 👍 or 👎 is one click and is counted as signal. They are the floor, not the ceiling.

Governance and Trust

Learn More

Related: CRM for the record memory agents read from, and Helpdesk for the same governance model applied to support.

Sign in to KarmaFlow