Chat & agents

The chat panel is how you operate Rounds as an AI SRE. Behind it is a multi-agent system: the orchestrator picks the right tool for each turn, the alert-rule agent handles alert lifecycle, and the investigation agent runs structured incident workflows across telemetry and (when configured) cluster state.

The agent talks SRE: incidents, metrics, logs, traces, dashboards, alerts, runbooks, remediations. Internally those map to "connectors" and "tools" — but you don't need that vocabulary to use the product. Connector terminology only shows up where it actually matters: settings, capabilities, and the auto-remediation reference docs.

What you can do

• Ask anything — "what dashboards do I have", "create one for HTTP latency", "investigate the 5xx spike at 9am"
• Operate dashboards by conversation — create, explain, clone, modify, rearrange, or delete dashboards without hand-editing JSON
• Investigate production symptoms — ask why latency, errors, saturation, or alerts changed; the agent can query metrics, logs, changes, and Kubernetes when configured
• Approve risky actions in-chat — read-only tools run inline; mutating actions surface as Run / Confirm / Apply with a risk note, evidence, and a one-line summary of what will change. No formal ApprovalRequest is created unless a permission gate or the GuardedAction risk model demands one. (Background-agent runs use a separate Approve / Reject / Modify path — see Auto-remediation.)
• Stream the agent's thinking — every step (tool call, result, decision) renders live in the panel as it happens
• Multi-tool turns — the agent can run several tools in parallel in one turn (e.g. 4 quantile queries at once) for speed
• Continue across sessions — chats persist; reopen a thread to continue
• Override the model — switch between Claude / GPT / Gemini / Ollama on a per-session basis
• Adjust effort — OPENOBS_THINKING_EFFORT=low|medium|high tunes how much extended thinking the model does before acting

How it works

Rounds implements a ReAct loop (Reason + Act). On each turn:

1. Model receives conversation history + tool definitions
2. Model emits one or more tool_use calls (native function-calling, no prose JSON parsing)
3. Rounds executes each tool, returns the result as tool_result
4. Loop repeats until the model emits reply / finish / ask_user

Because tool calls are native (not prompted JSON), the model picks tools more accurately and you can swap in any provider that supports function calling.

Available tools

Category	Tools	Used for
Discovery	datasources.list, dashboard.list, alert_rule.list, investigation.list	Find what exists
Metrics	metrics.query, range_query, labels, label_values, series, metadata, metric_names, validate	Discover + query metric backends
Logs	logs.query, logs.labels, logs.label_values	Search log backends
Kubernetes	ops.run_command	Inspect cluster state and prepare approval-gated remediation
Changes	changes.list_recent	Correlate with deployments / config changes
Web	web.search	External research (best practices, error codes)
Dashboards	dashboard.create, add_panels, modify_panel, remove_panels, add_variable, set_title, rearrange	Build / edit dashboards
Alerts	create_alert_rule, modify_alert_rule, delete_alert_rule, alert_rule.history	Manage alert lifecycle
Investigations	investigation.create, add_section, complete	Structured incident analysis
Navigation	navigate	Open a URL in the user's browser
Conversation	reply, finish, ask_user	Terminal actions

How to use it

Start a chat

Click the chat button (bottom right). Type a prompt. Watch the step trace as the agent works.

Continue a chat

Sidebar → Chats → click any past conversation. The agent loads the full history; you can pick up where you left off.

Switch models

Settings → Models → pick provider + model. Effective immediately for new chats. Existing chats keep their original model unless you explicitly retry.

Tune the effort

For complex tasks, more deliberation pays off:

export OPENOBS_THINKING_EFFORT=high   # low | medium (default) | high

low ≈ 1024 thinking tokens, medium ≈ 4096, high ≈ 16384. Only models that support extended thinking honor this — the rest silently ignore it. See supported models.

Cancel a long-running task

Click the stop button next to "Working..." in the chat panel. The agent receives a cancellation signal and emits a reply with what it managed to accomplish before stopping.

Supported models

Provider	Tool-use	Extended thinking
Anthropic	All Claude 3.x and 4.x	Claude 3.7+ and all 4.x
OpenAI	All GPT-4+ and o-series	o1, o3, o4, gpt-5.x (reasoning_effort)
Gemini	1.5+	2.5+ (thinkingConfig)
Ollama	Tool-capable models only (llama3.1, qwen2.5, mistral, qwen3)	Model-dependent — best-effort, not surfaced

Models without tool-use support are rejected at session start with an error pointing to a tool-capable model.

Permissions

The chat respects RBAC. The agent only sees tools the current user has permission to invoke:

• chat:use — required to open a chat at all
• Per-tool permissions — e.g. dashboards:write is required for dashboard.create; without it, the tool is removed from the agent's available set
• Folder-scoped permissions cascade to the tools that touch that folder's resources

So a Viewer can ask the agent to read dashboards but can't ask it to create one — the agent itself doesn't know about the missing tool, it just won't be in the prompt.

Limits

• Per-session token budget — large conversations get auto-compacted (older messages summarized) to stay under the model's context window. Crucial IDs (dashboardId, investigationId) are preserved in the summary.
• Multi-tool parallelism caps at the model's native limit (Anthropic: ~10, OpenAI: similar). The agent doesn't artificially cap.
• The agent is read-only by default for sub-tasks like investigations; mutations require the orchestrator path, RBAC, and approval where the target is an infrastructure change.

Related

• Dashboards — main use case
• Investigations — structured incident workflow
• Alert rules — alert lifecycle via chat
• Permissions — chat:use, agent:use, tool-level gates