Agent improvements

When a conversation goes wrong, you flag it. Infynia periodically reviews the conversations you flagged and proposes concrete, ready-to-apply edits to your AI agent’s configuration — a new rule, a wording tweak, or a behavior change.

For each suggestion you see a plain before/after comparison, the exact wording that would change, and the conversations that motivated it. You accept (optionally editing the text first) or reject. Anything you apply can be reverted in one click, and Infynia tracks whether each applied change actually improved your results.

Agent improvements are available on more complete plans — see plans and billing.

Where to find it

• From an agent card, on the AI Agent page, click Improvements.
• When there are pending suggestions, a dot appears on the AI Agent item in the side menu, and a notice with a shortcut to the suggestions appears when you edit the agent.

Flagging happens in the Inbox, via the Conversation feedback button.

How to use it

1. Flag bad conversations — in the Inbox, open a conversation that went wrong and use the Conversation feedback button. Pick one or more reasons (didn’t understand the customer, off-topic, wrong tone, repeated itself, wrong info, took too long, or “Other” with a note). You can edit or remove feedback later.
2. Give consent once — the first time you open Improvements, you must agree to a privacy notice before any analysis runs. You can revoke consent anytime, which stops future analysis.
3. Wait for suggestions — analysis runs automatically about once a day (end of day, Brasília time) on the conversations you flagged. New suggestions appear under Pending.
4. Review a suggestion — each card shows which part of the agent it targets, the change type (add / replace / remove), a title, and an explanation. Expand the preview to see the side-by-side before/after (additions in green, removals struck through in red). Expand the conversations list to open the exact chats that drove it.
5. Accept or reject — edit the suggested text in place if you like, then click Accept to apply it to your agent, or Reject to dismiss it.
6. Track impact & revert — applied suggestions show their effect once enough data is collected, and a summary aggregates how applied changes moved your results. Use Revert to roll any applied change back. A History tab shows accepted, rejected, reverted, and expired items.
7. (Optional) run it now — instead of waiting for the daily cycle, use Run analysis now (limited to once a day).

Note

Privacy. Consent is required: without it, nothing is analyzed. Analysis only looks at the conversations you flagged and sends a limited, truncated slice of recent messages (including customer messages) to an AI provider solely to generate suggestions — never to train third-party models. If you revoke consent, analysis stops.

Rules & limits

• Consent is required — without it, nothing is analyzed.
• It’s feedback-driven — no new flags means there’s nothing to analyze.
• It runs about once a day, and Run analysis now is capped at once a day.
• You can’t start a new run while one is in progress or while you still have unreviewed pending suggestions.
• Every suggestion is reviewed by you — nothing is applied automatically.
• Suggestions expire if left unreviewed for a long time.
• A revert can be blocked if you edited that part of the agent by hand after applying the change (so your manual edits aren’t lost).
• Impact results appear only after about a week of data.

Tip

The more problem conversations you flag with clear reasons, the better and more useful the suggestions tend to be.