Bot improvements

Your bot learns from its own conversations: the system collects signals from what goes wrong (execution errors, turns that run out of steps, fallback replies, human handoffs) and, when there is enough volume, proposes concrete adjustments to the agent. You always approve before anything changes.

How it works (3 steps)

1. Signals — the system automatically counts friction signals from recent days: tool errors, turns that exhausted their step budget, generic fallback replies and human handoffs. With at least 3 signals in the window, analysis becomes available.
2. Analysis — when you press Analyze, a model reviews those signals together with the conversations and generates improvement proposals (for example, clarifying the agent's instructions or tuning a flow's replies).
3. Approval — you review each proposal with its explanation and choose to Apply or discard it. Applying creates a new version of the bot (with a changelog) — history and rollback remain intact.

Where to find them

Open your bot and go to Improvements. There you'll see the active proposals with their status (proposed / applied / discarded) and the Analyze button to run a new round.

What improvements don't do

• Nothing changes without your approval — the flow is signals → analyze → approve.
• Nothing is lost on re-syncs — applied adjustments survive site re-syncs (they are re-merged on top of the new version).
• They don't touch your business data — only the bot's instructions, flows and replies.

::note
Best practice: run Analyze every week or two on bots with real traffic. Proposals are grounded in actual conversations — no traffic, no signals.
::