Agent Conversations
Agent Conversations is session replay for AI agents. When an AI agent uses your site’s agent tools, each visit is recorded in your workspace as one conversation: what it asked, exactly what your site answered, whether that was a real answer, and how far it got.
Your sales team gets it on each lead, as What their agent asked first and a link to Replay the conversation, so the first call starts where the research ended. Your marketing team sees every question agents ask your site, how often, and which ones it could not answer, including from buyers who never became leads. For what that is worth to each team, see Agent Conversations. This page covers setting it up and reading it.
Turning it on
Recording is on from day one for every new workspace. Once your site’s agent tools are live, there is nothing else to install and nothing to switch on.
There is no switch for recording, and no control for how long conversations are kept, in the dashboard yet. A workspace created before recording became the default (2026-09-24) keeps the setting it had; if yours shows no conversations after agents have used your tools, contact us.
What is recorded
A conversation is one visit by one agent, whichever of your tools it used. The dashboard labels where each question was asked:
| Where it was asked | What it is |
|---|---|
| MCP | Your hosted MCP endpoint (https://your-domain.com/mcp), called by an agent’s MCP client, often from its own servers |
| Agent endpoints | Your /agent/* answers |
| On your page | Your Agent Actions, which your pages register in the browser |
Calls from the same agent in the same sitting are joined into one conversation across these surfaces. A conversation ends after five minutes without a request, after 30 minutes, or after 100 requests, whichever comes first; the agent’s next request starts a new one.
Each conversation holds:
- The agent, as it identified itself. A client the product does not recognise shows as MCP Client.
- Each turn: the tool or endpoint it called, its question with contact details masked, and exactly what your site answered.
- The answer status of each question (below).
- How far it got: the furthest stage it reached.
- What came of it: its outcome, and the lead, if it produced one.
An agent that only fetches your pages can appear in the list as Looked around without asking: there is no question to read.
Answer status
| Status | What it means |
|---|---|
| Answered | Your site had a real answer for this. |
| Partly answered | A general answer went out for a specific question. |
| No answer | Your site had nothing to say. |
| Declined | A request your site chose not to answer. Not a content gap. |
| Error | Something broke while answering. An operations problem, not a content gap. |
A gap is a question marked Partly answered or No answer, nothing else.
How far it got
| Stage | What it means |
|---|---|
| Found | The agent discovered what your site offers. |
| Understood | It asked a question and your site answered. |
| Chosen | It weighed you up: pricing, fit, availability or a comparison. |
| Acted | It started a request. |
| Delivered | The request was confirmed. |
What came of it
| Outcome | What it means |
|---|---|
| Converted | It ended in a lead. |
| Stopped at confirm | The agent reached the confirmation step and the person did not say yes. |
| Informed | The agent got real answers and left without acting. |
| Unserved | Nothing it asked got a real answer. |
| Still open | Recent enough that it may still continue. |
Where to read it
In the dashboard, open Agent Intelligence → Conversations. It has three tabs.
Conversations. Every conversation, newest first, under the funnel How far your agent conversations get. Filter by where it was asked, by outcome, or to the conversations that Hit a question you could not answer, and search what agents asked. Open any conversation to replay it turn by turn. A two-sentence summary is written the first time you ask for one, not before.
Questions. What AI agents ask your site, and whether it had a real answer. Each row is a question: how often it was asked and in how many conversations, how each ask ended, and how often it went without a real answer. Sort by Most asked, Most unanswered or Most recent, and open a question to see the conversations that asked it. Questions are grouped by their exact wording for now, so two phrasings of one question are two rows.
Paths. The routes agent conversations take through your answers, what comes of each, and how each agent researches your site. A path is the first 12 steps of a conversation, with repeats in a row counted once.
A rate always travels with its counts (“9 of 34 · 26%”). Under 20, the percentage is withheld and marked too few to call.
What reaches your team with each lead
When a request is completed on your site, the lead arrives in your Leads inbox and by email. After the lead’s own layers (the person’s own words, the facts, how they approved, and which agent and surface it came from), both show What their agent asked first: up to five of the questions the agent asked before the request, each marked answered, partly answered or no answer, and a link to Replay the conversation in your dashboard.
- If their agent went straight to the request, the inbox says Its agent went straight to the request without asking anything first.
- If the conversation has not been linked to the lead yet, the email carries See the agent conversation behind this lead instead.
- The criminal-defense pack keeps the questions out of the email and shows the link alone. The dashboard still has them.
The email quotes the agent’s questions, not your site’s answers. The replay has both.
How the person said yes
Every request records how it was approved, as the lead’s approval_method:
| Recorded as | What happened |
|---|---|
card_tap | The person confirmed on the card in your brand, on your page. |
agent_authorized | Their own agent completed the request for them, with their details and their go-ahead. |
express_form | The person sent your express form page. |
An agent can complete a request, and you can answer it, but it can’t opt anyone into your marketing. Only the person’s own tick on the card does that. One contact can complete five requests a day with you, and agents that call too fast are told to wait. The full approval flow is in Agent Actions.
Ask your own Claude
Add the Inception Agents connector in Claude, then ask in plain words: What came in today? or What did agents ask us this week that our site could not answer? Claude answers from these tools:
| Tool | What it answers |
|---|---|
get_conversation_brief | Where to start: how agents used your site over a period, how far they got, and what they could not get answered |
list_conversations | Conversations, newest first, filtered by agent, surface, stage, outcome or gap, or searched by what was asked |
get_conversation | One conversation, replayed turn by turn |
get_conversation_questions | What agents ask your site, how often, and how each ask ended |
get_unanswered_questions | The questions your site had no real answer for, worst first |
get_conversation_paths | The routes conversations take through your answers, and what comes of each |
get_conversation_funnel | How far agent conversations get: found → understood → chosen → acted → delivered |
get_agent_research_styles | How each agent researches your site |
get_lead | One lead, by its reference |
record_lead_outcome | What came of a lead, and what it was worth |
Every tool above only reads, except record_lead_outcome. For a brief, run the /conversation-insights prompt: it turns the record into one brief with a lens each for marketing, product, sales and leadership. The sales lens lists, for each lead, the two or three questions its agent asked first. Every number in it keeps its denominator.
See it before you go live
Going live runs from Claude with the /inception-go-live prompt, and its last check is a live test. We send a test lead through your real site the way a buyer’s agent would: it asks a few questions, then requests a callback. The email your team gets is marked TEST, shows what the agent asked first, and links to a replay of the conversation. One of your people confirms it arrived (I received this test lead) before your site is marked live. Test leads are never billed.
Privacy and access
- Masking. Email addresses and phone numbers in an agent’s question are masked wherever they appear. Names, street addresses and secrets such as passwords, tokens and card numbers are masked when the agent sends them as named fields (a
nameorphoneargument, for example). - Retention. The words of each question and answer are kept 365 days. After that a conversation keeps its counts and stages, not its words.
- Erasure. An erasure request removes the conversations a person’s visits took part in. To make one, contact us.
- Visiting agents. They see your public tools only. Your conversations and your analytics answer only to your workspace.
- Who can read it. Anyone signed in to your dashboard, and any API key for your workspace, can read your conversations. Treat your keys accordingly.
Limits today
- It records conversations with your site’s agent tools. An agent that only reads your pages leaves no questions to read.
- Research from separate sittings is not joined. If their agent comes back another day, that is a new conversation, and the lead shows the conversation its request came from.
- What their agent asked reaches your lead email, your Leads inbox, your dashboard and your own Claude. Lead webhooks and CRM connectors do not carry it.
- Questions are grouped by their exact wording.
- There is no switch for recording and no retention control in the dashboard.