Production throws an error. Someone has to notice it, reproduce it, find the code responsible, understand why it broke, and write the fix. That is an hour on a good error and a lost afternoon on a bad one, and on Giant Context that someone is only ever me. Errors do not pile up here anymore, because the platform investigates them before I do.
Error triage is expensive because most of it is not the fix. It is everything before the fix. The reproduce, the search through the codebase for the line that threw, the reconstruction of what the code was trying to do, the realization of what it should have done instead. By the time I know the fix I have spent the hour, and writing the change is the last five minutes of it.
That hour has to come out of something. On a team it comes out of a person whose job it is. Alone it comes out of whatever I was building that day, so the cost of triage gets paid in features that did not get written.
Every error in production flows through the same event pipeline that carries everything else the platform records. When one arrives, a service I call Brain does the hour of work that used to be mine.
It files the error as a real issue, already triaged, with the priority, the occurrence count, the affected users, the stack trace, and a signature so duplicates collapse into one. Then it investigates. Brain is an agent with access to the codebase. It resolves the stack trace back to the file that threw, reads its way out from there through the code that called it, and assembles an explanation.
What lands is not something is wrong near here. It is a diagnosis. A recent one, on a route that renders a site's robots file, read like this. The code queried a table named domain that does not exist. The correct table is project_domains in the core schema, and the column is hostname. And because a diagnosis alone would still leave me the work, Brain wrote the change:
The fix Brain wrote
- const domainResult = await websiteDb.query(- `SELECT d.domain- FROM domain d+ const domainResult = await coreDb.query(+ `SELECT d.hostname AS domain+ FROM project_domains d WHERE d.app_id = $1+ AND d.app_type = 'website' AND d.is_primary = true`,The report ends with the steps Brain took. The searches it ran and the files it opened, in order. It shows its work rather than handing me an answer to take on faith, so I can check the reasoning before applying anything.
The report is not where it stops. Brain runs an MCP server, the same kind of tool interface the platform exposes to customers, and I connect to it from Claude Code. I pull the open issues, read Brain's explanation against the code, and if it holds up I apply the fix Brain wrote and deploy. The investigation was Brain's. Mine is the judgment, whether the diagnosis is sane and the fix is right.
The hour became a review. An error I would have dreaded arrives with its cause named and its fix drafted, and my job shrinks to reading the reasoning, agreeing or correcting, and committing.
Working alone, this is the difference between a product whose errors are handled the day they appear and one where triage keeps losing to feature work. Now it mostly does itself.
I pointed the platform's own intelligence at errors before anything else, because errors are the problem where a wrong answer is obvious to me. They were my bugs. If a diagnosis were nonsense I would know in seconds. They have not been nonsense often.
There is an order to this, and each one needs the one before it.
Marketing that runs itself.
Software that heals itself.
Software that builds itself.
A platform that operates itself.
A company that needs no humans at all.
Remote, Pacific time, full-time or contract. Get in touch.