secure-agent-workflow
A secure software factory for coding agents. Agents build from approved issues with little supervision, and nothing reaches production unchecked.
The agent can't change the controls: they live in the sandbox, GitHub permissions and cloud credentials, not in instructions or settings files the agent can read and edit.
Overview: the one-page picture shows the topology, the workflow and how a project adopts it. It describes the full design, including the parts not built yet. Source: docs/index.html.
Status: these are built and tested end to end on testbed repos: the agent's GitHub App, bootstrap, the merge gate (checks, scanners, tier check), the AWS credential broker, daily credential checks, and Cloudflare previews and staging. This repo's own PRs go through the same gate. The orchestrator has its webhook intake, token minting and outbound policy, which is unit-tested but hasn't run against a real container yet. The sandbox, infrastructure as code, production deploys, PR audit and red-team runs are not built yet.
How it works
- A human writes the spec, plan and ADRs, and turns the plan into issues grouped in waves.
- Dispatching a wave labels its issues
agent:ready. Each one starts a build agent in a fresh, locked-down sandbox, in dependency order and within a concurrency cap and spend budget.
- The agent works on its own branch and opens a PR, which gets a preview deploy. It can't touch workflows, settings or
main.
- A merge gate routes the PR by what it touches. Low-risk changes merge on green checks; risky ones wait for a human.
main deploys to staging in the dev target. Production, in the prod target, is a versioned release PR from main into a production branch, approved by a human, and only that deploy can get production credentials. The release path is designed and gets built after the path to staging works.
Accounts
The platform and the apps it builds run in separate accounts:
| Term | What it is | Holds |
|---|
| Factory account | The platform's own cloud account. Never a deploy target | Orchestrator, sandboxes, operator console |
| Dev target | Where the app deploys previews and staging. One account per provider the app uses, such as a Cloudflare account, an Auth0 tenant or a database project | Preview and staging deploys |
| Prod target | Where the app deploys production, again one account per provider | Production deploys |
| Broker account | The key store. Its dev and prod stores can share one AWS account or use two | OIDC roles, Parameter Store, OpenTofu state |
A Cloudflare token that can deploy Workers can change every Worker in its account. That's why the factory account is never a deploy target: an app's deploy credentials can't reach the orchestrator.
Pieces
| Folder | What it holds | Status |
|---|
bootstrap/ | One command that applies rulesets, environments and gate wiring to a repo. Also the agent's GitHub App manifest and scripts | Built |
gate/ and .github/workflows/ | Reusable merge-gate workflows: checks, scanners, tier check, PR audit | Built, except the PR audit |
standards/ | ADR template, agent instructions, issue template, planning agent instructions, tier map | Built |
orchestrator/ | Cloudflare Worker that runs agents in sandboxes, with credentials kept outside | Webhook intake and token minting. No sandbox yet |
deploy/ | Deploy template with a GitHub OIDC credential broker | AWS broker, credential checks, and Cloudflare previews and staging built; OpenTofu and production not started |
redteam/ | Attacks the controls must stop | Not started |
Adapting it
v1 runs on Cloudflare and AWS. Each part lives in its own folder and states what it expects from the others, so you can rewrite one part without touching the rest. There is no configuration layer for this; you change the code.
| Part | v1 choice | Other options | Folder |
|---|
| Agent identity | GitHub App (ADR-001) | Machine user account; a GitHub org on Team, which adds push rulesets | bootstrap/ |
| Sandbox | Cloudflare containers (ADR-004) | E2B; Docker Sandboxes on your own host; containers under gVisor or Kata; GitHub Actions runners on public repos | orchestrator/ |
| Egress and credential proxy | Worker outbound handlers | Envoy or Squid with an allowlist in front of any container host | orchestrator/ |
| Agent CLI | Not yet chosen | Any agent CLI that runs headless, such as Claude Code or Codex | orchestrator/ |
| Scanners | Free tools; GitHub's own on public repos | GitHub Advanced Security, or a commercial scanner | gate/ |
| PR audit models | Builder and reviewer from different model families | Any model API, including a self-hosted model | gate/ |
| Credential broker | GitHub OIDC to AWS IAM and Parameter Store, with separate preview, staging and prod roles in one account or two (ADR-002) | Vault or OpenBao; GCP Workload Identity with Secret Manager; Azure federated credentials with Key Vault; GitHub environment secrets, which is simplest but long-lived | deploy/ |
| Deploy target | Cloudflare Workers | Any target the deploy job can reach with brokered credentials | deploy/ |
GitHub doesn't swap out. Bootstrap, the gate and the agent identity all depend on GitHub rulesets, Apps, Actions and OIDC.
Whatever you replace, keep the property the design depends on: the agent cannot change the control.
Related work
"Software factory" is the name in use for systems where agents build software from specs and issues, and people set direction and approve risk. These projects work in the same space. Descriptions are from their own sites and repos, November 2026.
여기까지가 README 앞부분입니다. 나머지는 원문에서 읽을 수 있습니다. README 전문 보기 ↗
