Investigations

An investigation is the agent engagement that works one confirmed issue. Each issue gets at most one: a coordinating agent embedded in the issue, whose conversation lives in the issue's Investigation tab. Under the hood it is a thread of type investigation, but you never leave the issue to follow it.

How investigations start

• Triage confirms an incident. When triage lands the Incident verdict on an issue, Polylane starts the investigation automatically, subject to the daily limit below.
• An automation runs the openIssue action. Any automation can open an issue with a title, a severity, and a prompt. The issue's investigation starts with it, subject to the same limit.
• You hand off from a chat. In any chat thread, the agent can call openIssue to turn the problem you are discussing into an issue and start its investigation immediately.
• You press Investigate. Every issue without a running investigation shows the button. Manual starts never count against the daily limit.

Because an issue has at most one investigation, a re-fire of the same problem reopens the existing issue and nudges the same investigation instead of starting a second one.

Held investigations and the daily limit

Workspaces cap how many investigations start automatically in any rolling 24 hours. Past the cap, the issue keeps its Incident verdict, is marked Investigation held, and shows a Start investigation button. Workspace admins are emailed, and an admin can start a held investigation at any time; doing so does not consume a slot.

Settings, Investigations controls two things:

• Parallel passes per hypothesis: 1 to 12, default 3. More passes are more thorough and use more credits.
• Auto-investigation limit: the cap on automatically started investigations in any rolling 24 hours. The Free plan starts at most 10 automatic investigations per day and the limit cannot be raised. Paid plans can set the limit or turn it off.

Verdicts are never limited. Every confirmed issue is visibly an incident on every plan; the limit only defers the investigation.

Hypotheses

The coordinating agent proposes hypotheses, up to 5 per run. Each hypothesis is tested by parallel passes: hypothesis sub-agent threads that each get an argumentative role (prosecution, defense, or neutral) and a different starting point, such as logs, metrics, recent changes, the topology, code, traces, or past issues. Each pass returns a verdict of confirmed, refuted, or inconclusive with a confidence level from definitive down to speculative. A majority vote across passes decides each hypothesis, and the run compiles a headline conclusion.

Outcomes and parking

After each agent turn, Polylane classifies the investigation:

• A diagnosis lands. resolved means root cause found and the problem is fixed or recovered. diagnosed means root cause found but the fix has not landed. inconclusive, false_positive, and failed cover the rest.
• The investigation parks. If the agent needs your action (needs_human_action), is awaiting your decision (needs_decision), is waiting for a fix to ship (awaiting_change, for example an autofix pull request awaiting review), or is stuck on an error it could not work around (failed), the investigation parks and monitoring stops nudging it to re-investigate. Parks that need a human are recorded as escalations, so the same ask never notifies you twice.

Timeline

Everything the investigation does lands on the issue timeline: investigation started or held, hypothesis results, and every autofix step from branch push to pull request merge, alongside detections, re-detections, triage, and your own notes and milestones.

How investigations close

Investigations close with their issue. When the issue resolves, on recovery or by hand, the investigation closes and any parked status and diagnosis clear, so a later re-detection reopens the same issue and takes a fresh look instead of trusting a stale conclusion.

Use polylane to list the issues under investigation in my workspace, and for the most recent one summarize each hypothesis, its verdict, and the confidence behind it.

Related

• Issues for the record every investigation is embedded in.
• Threads for how investigation and hypothesis threads relate to chat and autofix threads.
• Escalations for the ledger of parked asks that need you.
• Automations for opening issues automatically with openIssue.
• Autofix for turning a diagnosis into a pull request.