TofuPilot Agent

Last updated on May 21, 2026

PreviewThis feature is in preview, so its scope and behavior may change. Request access.

The TofuPilot Agent runs an autonomous root cause analysis (RCA) inside the chat. You point it at a failure pattern; it generates hypotheses, gathers evidence from your runs, and ranks the leaves of the hypothesis tree by statistical confidence.

When to use it

Reach for the agent when:

  • A procedure has dropped in yield and you do not know which phase or measurement is responsible.
  • A batch is failing at the same step and you want to confirm whether the cause is the part, the station, or the test itself.
  • A measurement is drifting and you need to localize which dimension (revision, station, operator, time) explains the drift.

The agent is not the right tool for one-off lookups; use chat directly for those.

How it works

The agent runs a guided loop:

  1. Frame. You describe the failure pattern in natural language ("FVT yield dropped from 95% to 78% this week").
  2. Hypothesize. The agent generates candidate causes (phase, measurement, station, revision, operator).
  3. Gather evidence. For each leaf hypothesis, the agent queries runs, measurements, and Cpk data through the same tools as the chat.
  4. Score. Each leaf is scored with a Wilson 95 confidence interval against a baseline.
  5. Rank and report. Leaves with the strongest signal surface at the top, with the evidence rendered inline.

The agent surfaces its working tree in the chat, so you can prune branches, add hypotheses, or jump to the dashboard to inspect raw data at any step.

Hypothesis statuses

Each leaf carries a status that reflects where the evidence stands.

StatusMeaning
pendingHypothesis generated, evidence not yet collected.
investigatingTool calls running; partial evidence visible.
supportedConfidence interval crosses the threshold. The hypothesis is a likely cause.
rejectedConfidence interval rules the hypothesis out.
inconclusiveNot enough data to score. Agent suggests what more data you would need.

What the agent can see

The agent runs through the chat backend, so:

  • Every tool call is scoped to your role, team, and SSO group mappings.
  • Stations and Viewers can run RCAs against the data they have access to.
  • Personally identifiable information (operator names, emails) is shown only when your role would expose it on the dashboard.

Saving the result

You can save the agent's final report as a chat message link, attach it to a ticket, or paste the hypothesis tree into a Report to share with your team.

How is this guide?

On this page