Intercom's development workspace cannot be converted into a production workspace, is available only in the US region, and keeps at most twenty users or leads, archiving the oldest ones automatically when the cap is passed (Intercom developer getting started). A test workspace on the higher plans is a separate workspace with separate data, and Intercom's help center says it is not possible to migrate data from a test workspace to the live one. So an agent that works conversations is tested in a workspace with none of its customers in it, capped at twenty of anyone, with no way to reset it between runs or to break it on purpose. The Intercom world is a stateful Intercom mock built for testing agents: the same URLs, the same errors and the same state machine as the real API, on your machine, verified against the real system.
Intercom test workspace limits
Twenty users is not a queue. The conversations that catch an agent are the ones with history: the customer who wrote three times about the same order, the lead who was already handed to sales, the thread another teammate snoozed for a reason. A test workspace has to have all of that typed into it, and after one run it holds the leftovers of that run instead. A world is a pseudonymized copy of your own workspace, taken when you ask for it: the same contacts and companies, the same conversations in the same states, the same teams and assignments, under new names. Every run starts from those bytes, in milliseconds, and ends by throwing the world away.
A stateful Intercom mock: conversations, assignments, snoozes, resolution flows
The world keeps contacts and companies with their attributes, conversations with their parts, replies and notes, assignments to admins and teams, tags, snoozes and the open, snoozed and closed states a conversation moves through. It keeps Intercom's rules with them. A closed conversation reopens when the customer writes again. A snoozed conversation returns at its time. A note is not a reply, and the customer never sees it. An assignment to a team is not an assignment to a person. Everything the agent creates or changes stays in the world, so the second request sees what the first one did, and the wrong sequence gets the real error in the real words.
Test your AI agent against Intercom
Point the agent's Intercom client at the world. Nothing in the agent changes. The path is three commands: the first copies the workspace and writes the worlds folder, the second turns your own conversation export into tasks, one per row, with the expected outcome on the rows where you know it, and the third runs the agent three times on fresh copies and writes the report. The three steps take about ten minutes; the GET ACCESS form on this page is where the Intercom world starts.
Grade what the agent did, not what it said
A run starts from a mark and ends with a diff: every reply sent and to whom, every note, every assignment, every snooze and its time, every conversation closed or reopened, every tag. The transcript says the conversation was handed to billing; the diff shows it still assigned to the agent with a note nobody will read. The transcript says the customer was told to expect a refund; the diff shows the reply, and the payments world beside this one shows whether any money moved, in dollars. The report gives every task a verdict and a reason, and replays the worst run to the byte.
Break it on purpose
Intercom rate limits its API, and an agent working a queue will meet the limit at the worst moment. In the world that moment is on your schedule: a 429 on every third call, an outage mid run, a slow API, an expired token, a webhook delivered twice. Does the retry logic send the same reply twice? Does the agent close the conversation when the reply never went out? Then move the clock, so snoozes return and the reopened conversations pile up, and read the diff again.
Intercom test environment for agents, in CI
The same command that runs on a laptop is the CI step the first command wrote. No workspace to keep under twenty users, no data to rebuild: the world starts fresh inside the job, the agent runs every task, the report is the artifact, and nothing leaves the machine. Every world in the catalog takes the same path; Zendesk, HubSpot and Stripe are three more of them, and the docs are the path in full.