Investigate
Ask the agent about incidents, latency, deploys or capacity, and follow every tool call it makes.
Investigate is where your team talks to the agent. Everyone except Viewers can use it. The header shows how many tools your role can use right now, and on which servers.
Ask a question
Describe what you're seeing, the way you would to a colleague:
- Summarize activity in the last 12 hours
- Any error spikes or failing endpoints today?
- Why is p99 latency up on checkout since the 14:00 deploy?
- Which pods restarted in the payments namespace this morning, and why?
Press Enter to send, Shift+Enter for a new line. Click Stop to end a run early.
What happens while it works
- Thinking… while the agent plans its next step.
- One line for its tool calls, showing the tool running now (for example github · list_commits). When they finish it reads Used 8 tools on github. Click it to see every call, then click a call to see its input and output.
- The answer streams in as it's written, usually in four parts: Summary, Evidence (with the tool behind each finding), Likely cause and Suggested next steps.
Under each answer, Copy copies it as Markdown, and Run again asks again. New investigation starts over.
What the agent can use
The agent only sees the approved tools of active servers that your role may use. A Developer whose role is granted only the staging cluster can't investigate production through sremcp, and the agent won't see production's tools in their chats. See Role access.
The agent can't change anything you didn't approve: when remediation is needed, it describes the exact steps for a human to take.
Limits
- A run takes at most about 4½ minutes, and each tool call at most 60 seconds.
- The agent takes at most 15 steps per answer.
- At most 64 tools are offered to the agent at once. With more approved tools than that, only the first 64 load, and the answer says so. Approve the tools you actually need.
- The whole conversation is sent with each question. Very long conversations are refused with Start a new one.
Is my chat saved?
The conversation on screen isn't: reloading the page or clicking New investigation clears it. But every answer is recorded in Runs, with the question, the answer and the tools it called. Managers and above can see everyone's runs.