Claude Code investigates Canvas support bugs
Gunderson reports that Canvas’s Claude Code Support Stack retrieved GitHub, Pylon and Sentry evidence, found a customer bug and prepared a fix he shipped; added context later corrected false plugin blame.
canvasmedical.com
Suggested prompt
Suggested prompt written by Agents In Use, not the source author. Replace bracketed inputs and review tool permissions first.
Investigate [customer question] with [Support Stack sources]. Trace symptoms through support history, code and error evidence. Distinguish correlation from a plausible mechanism, and prepare a repair for my review before shipment.
Reach people exploring AI agents
See placements →Evidence at a glance
Sourced account- Source author
- Beau Gunderson
- Original account
- Read the source ↗
- Published
- Source link checked
- Reported result
- Gunderson reports that Canvas’s Claude Code Support Stack retrieved GitHub, Pylon and Sentry evidence, found a customer bug and prepared a fix he shipped; added context later corrected false plugin blame.
- Human involvement
- Gunderson accepts and ships the proposed fix. The team corrects false plugin blame through context updates; company time savings are not independently measured.
A linked account supports attribution. It does not mean we independently reproduced its result.
Report an issue with this case
Open correction emailOpens your email app with the case link included. Nothing is sent automatically.
How the agent did it
- Clone the role-specific MCP and investigation-skill package.
- Provide the customer question and retrieve issue, code and error evidence.
- Inspect the bug finding and proposed repair.
Beau Gunderson and colleague Jess package Canvas's MCP definitions and investigation skills into a repository staff can clone. The resulting Support Stack uses Claude Code with company-specific context rather than asking support workers to carry every tool's history in their heads.
In the first reported test, a customer has just asked a question. Support Stack retrieves evidence from GitHub, Pylon and Sentry, identifies a bug and offers to draft a fix. Gunderson accepts that step and ships the fix to the customer. Staff then adopt the internal tool and the team iterates on its instructions when findings are confusing or unhelpful.
One failure is specific: it blames a customer plugin for instance-performance trouble even though that plugin cannot cause the observed performance effect. The team adds context to distinguish temporal correlation from a plausible cause. The source's later Slack Investigator uses the Claude Agent SDK, but is a separate evolution; this entry does not attribute every later automated action to the original terminal tool. The useful deployment pattern is a role-specific investigation package, live source retrieval and a human decision to ship a proposed repair. Time savings are company reports, not an independently controlled benchmark.
Prompt
Suggested prompt (written by Agents In Use, not the author's)
Source: Blog · Beau Gunderson
Frequently asked questions
Which evidence sources were used in the first test?
The source names GitHub, Pylon and Sentry in the investigation that found the customer bug.
Who shipped the repair?
Gunderson accepted the agent’s offer to draft a fix and shipped it to the customer.
What mistake prompted a context update?
The agent blamed a customer plugin for performance trouble despite no plausible mechanism connecting that plugin to the effect.
More details
Remaining steps
- Have Gunderson accept and ship the fix to the customer.
- Add missing context when a later investigation makes an implausible causal claim.
About this account
Claude Code retrieves support and engineering evidence through the Support Stack MCP package, investigates the bug and drafts a repair before human shipment.
Firsthand account by Beau Gunderson; original publication date verified against Primary Canvas Medical byline April 16, 2026; public JSON-LD datePublished 2026-04-16. Outcomes are the author’s reports, not independently reproduced. Suggested prompt written by Agents In Use; not the author’s original prompt.
Credit: Beau Gunderson
Source published
Added to Agents In Use
Source checked
Tools and topics
Claude Code · Support Stack · GitHub MCP · Pylon MCP · Sentry
Customer bug investigation · Human-shipped fix · Causal-context correction
Editorial completeness: 100/100. Measures documented fields, not agent performance or verified outcomes.