Prompts
Use Prompts to see what your agents are asked to do: the prompts people typed and the tasks one agent handed to a subagent, with identical text grouped as one entry. Each entry links to the sessions it ran in, where you read the actions that followed and the verdict each one received. Every role can read it.
| Where | What it lists |
|---|---|
| Prompts, in an agent’s Artifacts tabs | That agent’s prompts |
| Inventory, Prompts tab | Every agent’s prompts, one row per agent and prompt, each naming its agent |
The command palette (Ctrl+K) also finds prompts by their text. Each result names its agent and opens that prompt.
Find a prompt and what it led to
Section titled “Find a prompt and what it led to”- Open an agent’s Prompts tab, or the Prompts tab of Inventory.
- Search by prompt text, or pick a date range. In the inventory, narrow the list to some agents with the Agents filter.
- Select a row to open the prompt: its full text, which you can copy, its detail, and the sessions it ran in, newest first.
- Open a session to read the actions that followed the prompt and the verdict each one received. Last run opens the newest of those sessions in the agent’s Sessions tab.
The figures at the top count what your filters select: Prompts (entries), Runs, Sessions that carried a prompt, and Subagent tasks, the share of runs that were tasks handed to a subagent.
| Column | What it shows |
|---|---|
| Prompt | The text of the newest run, on one line |
| Agent | In the inventory only: the agent that ran it |
| Origin | Typed, Subagent task, or both |
| Surfaces | What the actions after each run reached, up to the next prompt of the same session: Local files, Shell, Web or MCP servers, with a count for each. A dash means no action followed |
| Sessions | How many sessions the prompt ran in. Select the number to open the prompt on its list of sessions |
| Run count | How many times it ran |
| Last run | When it last ran |
What is captured
Section titled “What is captured”| Origin | What it is |
|---|---|
| Typed | A prompt a person typed into a covered agent |
| Subagent task | The text an agent handed to a subagent as its task |
Subagent tasks come from these agents:
| Agent | Subagent tasks |
|---|---|
| Claude Code | Tasks handed to teammates, and to subagents that run inside the lead session. A task of the second kind counts under the lead session |
| Cursor | Tasks handed to subagents |
| Cline | Tasks handed to subagents |
| Codex CLI | None: a Codex CLI subagent’s task never reaches OpenLatch |
Prompts whose text is identical once whitespace is normalized are one entry. Runs of spaces, tabs and line breaks count as a single space, leading and trailing whitespace is ignored, and equivalent Unicode forms are treated as one. Case and punctuation still count, so “Fix the build” and “fix the build” are two entries. An entry shows the text of its newest run.
OpenLatch keeps the first 4,096 characters of a prompt. Two prompts that match up to that point and differ after it stay two entries, even though they show the same text.
How long prompts stay
Section titled “How long prompts stay”Prompts covers the last 30 days. A date range starts no earlier than 30 days ago; a range picked further back shows the last 30 days. Each day, prompts older than 30 days leave the tab. Deleting a session or an agent removes its prompts.
The 30 days apply to this tab only. A prompt is also part of the agent’s event record, which follows its own retention.
Who can see prompts
Section titled “Who can see prompts”Every role, from viewer to owner, reads the full text of every prompt in the organization. Capture is always on, with no opt-out: no setting or role turns it off.
Prompt text is shown as plain text. Markup or code inside a prompt is displayed as written, never run or rendered.
Redaction and your responsibilities
Section titled “Redaction and your responsibilities”Before a prompt is stored, OpenLatch masks text that looks like an access token or an email address. Anything it does not recognize is stored and shown as received: a key pasted in a format the redaction misses stays readable on this tab, to every role, for up to 30 days. Rotate any key that someone pasted into a prompt.
Let your employees know that prompts sent to covered agents are recorded and visible to authorized users of your OpenLatch console.
To read what an agent did after a prompt, see Investigate what an agent did. To find other things across every agent, see Search the inventory.