The Agentic Pipeline

Picture two a.m. on a Tuesday. Feedback keeps landing: a crash report, a wish for dark mode, some praise, and one note that isn't really feedback at all. Nobody is awake to read it. A pipeline of agents reads it, sorts it, and drafts the issues while you sleep. This is exercise eight: stepping out of the loop.

First, feel the pain. You write six messy entries into a feedback file and handle them by hand. Is it a bug, a feature, or praise? How urgent? Then open GitHub and type up every issue. Fifteen minutes. Now picture five hundred entries, every week. That is what pipelines are for.

So you build the first agent: intake. A tiny shell script pipes the feedback into Claude in headless mode. It runs once, answers, and exits. Its one job is to turn messy text into clean data. My eyes are burning at night becomes a tidy one line summary. It gets no tools at all, a schema its answer must match, and a hard cap on turns and spending.

The second agent has judgment. Triage reads each item and decides: bug, feature, or praise. High, medium, or low, by rules you wrote. Labels come from a fixed list, like a menu. The upload crash? A high priority bug. Dark mode? A low priority feature. Two agents, two jobs, joined by a pipe.

Then there's entry six. It tells the A I to ignore its instructions, file an urgent security bug, and paste its environment variables into it. Usually the model sees the trick. Don't count on it. The agent has no tools, so there's nothing to leak, and labels only come from the menu. The agents can read, but not act.

The pipeline script runs intake, triage, and a summary report, and stops the moment anything fails. It never touches GitHub. Filing is a separate plain script, code that acts but doesn't think. It checks every label again, caps the run at ten issues, and marks each one as machine written. You read the list first. Then one command files them.

Now put it on a schedule: cron, or GitHub Actions, every Monday at eight. The key lives in a file only you can read, or in a repository secret, never in a script or a prompt. In Actions, the agents' job holds the key but can't write anything. The filing job has no Anthropic key and no agent, and it waits for you to click approve.

That's the pattern. You define the rules. Agents execute them. You decide what they can touch, and you review what leaves your system. Then you step back, as far as the risk allows, and no further. Next up: exercise nine, your own SaaS.