We Built Our Own PR Agents, and You Can Too

Why It Matters

The approach lets engineering teams automate repetitive PR tasks while preserving human oversight, accelerating delivery and reducing errors across the software lifecycle.

Key Takeaways

• Bugbot pattern: separate branch, one‑click merge button
• Skill‑agnostic pipeline runs any custom agent
• Cloudflare Workflows ensure durable, retryable execution
• Typed skill output enables automated reviews without parsing prose
• Sandboxed execution protects secrets and limits network access

Pulse Analysis

AI‑driven code reviewers have exploded onto the market, but most services focus on generic linting or bug detection and charge premium fees. What organizations truly need is a way to embed domain‑specific knowledge—such as analytics standards, security policies, or API documentation rules—directly into the PR flow. By abstracting the agent’s logic into a reusable "skill" file, developers can swap out functionality without rewriting the surrounding infrastructure, turning a single CI‑style pipeline into a multi‑purpose automation hub.

The technical backbone leverages Cloudflare’s edge platform. Workers receive GitHub webhooks and launch durable Workflows that orchestrate isolated Sandboxes for each PR. Inside the sandbox, the Flue framework provisions scoped proxies for Anthropic and GitHub APIs, ensuring the LLM never sees raw tokens. OpenCode executes the skill, returning structured Valibot‑validated data that the workflow can immediately translate into inline comments, diff previews, or merge actions. This typed output eliminates fragile text parsing and enables deterministic automation, while the two‑phase prepare‑publish model guarantees that no code lands without explicit human approval.

From a business perspective, the pattern reduces manual effort, shortens review cycles, and enforces internal standards at scale. Teams can roll out new skills—like auto‑generating test suites or synchronizing OpenAPI specs—by simply adding a markdown definition, instantly extending the automation surface. Because the system is built on serverless, pay‑as‑you‑go services, cost scales with usage, making it attractive for both startups and enterprises seeking to embed AI safely into their development pipelines. As more organizations adopt skill‑agnostic PR agents, the competitive edge will shift toward those that can rapidly prototype and enforce bespoke engineering policies without sacrificing security or auditability.