# We Are Moving Our Agent Pipeline from GitHub to Cloudflare

Our coding agents used up a month of GitHub Actions minutes in 9 days, the third month in a row they ran out. Issues had no fields for the data agents store, and agents had no index of what the issues meant. The plan for rebuilding the pipeline on Cloudflare D1, Vectorize, Artifacts and Workers Builds.

Date: 2026-11-16
Author: Patrick Heneise
Source: https://zentered-studios.com/blog/github-to-cloudflare/

---

![An outlined arrow from a stack of cards to a cloud](/blog/github-to-cloudflare/cover.svg)

GitHub stopped starting our CI jobs. The annotation on each run read: "The job was not started because recent account payments have failed or your spending limit needs to be increased." No payment had failed. Our agents had used up the month's included Actions minutes in 9 days, the third month in a row they ran out.

A higher spending limit would have cost little. We started a new repository the next day anyway, because the tracker under the pipeline had two problems that more minutes would not fix. This post is the plan for the new pipeline. Six more posts follow over the next three weeks, one per part.

## What the old pipeline was

Agents picked up GitHub issues, opened pull requests, waited for Actions and merged. We wrote about that setup in [Agent-Driven Development](/blog/agent-driven-development-roadbook/).

Each issue first moved across a GitHub Projects board, from backlog to ready to in progress to done. Then we dropped the board for Milestones: a milestone names an outcome, and its issues are open or closed. We measured that switch in [the agent loop's setup cost](/blog/agent-loop-setup-cost/).

Milestones did not fix two problems with issues.

An issue has a title, a body and comments. An agent also needs the owner's intent, a definition of done, the command that proves the result, the size of the job and the model that should build it. GitHub has no field for any of these. They went into the issue body as formatted text, and every program that read them had to parse that text back out.

Agents had no index of what the issues meant. A product owner or product manager can keep a backlog somewhat sorted in their head, probably up to about 100 GitHub issues. An agent starts every session with nothing sorted. To find a duplicate, a related issue or the next thing that was ready, a session listed the issues and read them, every time. Any analysis across the issues started the same way.

This move is the third step. It puts direction, Tasks and knowledge in one database and indexes them by meaning.

## The plan

| On GitHub | In the new pipeline |
| --- | --- |
| Issue | A Task, one row in a [Cloudflare D1](https://developers.cloudflare.com/d1/) database per product |
| Label that marks state | A `status` column, changed by one SQL statement |
| Milestone | A row with a title, a goal and a target date |
| Decision in an issue comment | A row in a `knowledge` table |
| Search across issues and comments | Keyword search in D1 and a [Vectorize](https://developers.cloudflare.com/vectorize/) index per product, behind one recall command |
| Duplicate and related issues | Tasks found by meaning in that index, and links between them (`duplicate_of`, `relates`, `blocks`, `parent`) |
| Pull request | A fork per Task in [Cloudflare Artifacts](https://developers.cloudflare.com/artifacts/), merged into a `next` branch |
| Review on the pull request | A review row, written by a second model |
| Checks in Actions | The same commands, run on our own machines before the merge |
| Build and deploy | [Workers Builds](https://developers.cloudflare.com/workers/ci-cd/builds/), connected to the Artifacts repository |
| Notification | `factory inbox`, a command the owner runs |
| User feedback filed as an issue | A [Workers AI](https://developers.cloudflare.com/workers-ai/) gate that screens the text before it becomes a Task |

One Worker and one command line tool hold this together. Both are called `factory`. Agents never talk to D1 directly. They call the CLI, the CLI calls the Worker, and the Worker checks a short-lived token on every route. In Artifacts, a build session reaches only its own Task's fork.

The agents themselves do not run on Cloudflare. Build sessions are headless Claude Code on our own machines, on a Claude subscription. Reviews go to a Codex model. Cloudflare's documentation says [Claude Code behind its AI Gateway](https://developers.cloudflare.com/ai-gateway/integrations/coding-agents/claude-code/) is billed through Cloudflare credits or an Anthropic API key. That is a different cost from the subscription we already pay for.

## The index agents were missing

Vectorize gives agents the sorted backlog that a product owner keeps in their head. Each product has one index. A [Workers AI](https://developers.cloudflare.com/workers-ai/) model embeds every milestone, Task, decision and lesson when it is written. An agent asks in its own words, and recall returns the records that mean the same thing, even when they share no word with the question.

The D1 table stays the record, with ids, status and the owner's exact words. The index is a second way to find a row, and one command rebuilds it from the table.

The analysis that meant reading every issue is now a SQL query, because a Task is a row.

## Where Cloudflare's models fit

Besides the embeddings, Cloudflare's own models get two jobs. The first is the feedback gate in the table. The second is a nightly job that reads the day's failed checks and returned Tasks and proposes lessons from them. The owner accepts a lesson before any agent reads it.

## What stays on GitHub

GitHub is still the place for open source. Our open-source projects stay there, [mcpsignals](https://github.com/zentered-studios/mcpsignals) and [rivian-mcp](https://github.com/PatrickHeneise/rivian-mcp). Outside contributors need issues and pull requests, and Actions on public repositories is free on standard runners.

## The risk

All of it is new and little of it is tested.

Cloudflare Artifacts is in [open beta](https://developers.cloudflare.com/changelog/post/2026-10-01-artifacts-open-beta/). Its [pricing page](https://developers.cloudflare.com/artifacts/platform/pricing/) says billing for operations and storage starts this month, with 10,000 operations and 1 GB included per month.

The documentation is young. We know of nobody else running an agent pipeline on Artifacts, so there is no setup to copy and nobody to ask.

## The series

The next posts cover the Task table, the knowledge search, a repository per Task, CI minutes, the Cloudflare models we plan to use, and how we judge whether the pipeline works.
