TechPrufer
Get started
Cloudflare
Cloudflare logoCloudflare · Feature trail

AI-assisted Issue Triage and Software Factories

4 posts · Jun 2025 – Oct 2026

What this trail covers

Cloudflare OS, an open-source agent workspace, is now available as a fully managed deployment option. New capabilities include mounting existing Git repositories for agents to work with code, improved integration with Google Workspace (Gmail, Drive, Sheets, Calendar) via a Gatekeeper service, and enhanced export options for documents, presentations, and spreadsheets to common formats like .xlsx, CSV, PDF, Markdown. The AI agent harness for security operations speeds up data gathering, aggregation of detections, and accounting for missing sources, using specialist AI agents for traffic analysis, customer context, global telemetry, and threat intelligence. It leverages Clef for triage and decision-making, Workers AI for agentic reasoning, Workflows for orchestration, D1 for state persistence, R2 for evidence storage, and Durable Objects for case-chat state.

All Cloudflare trails
  1. Building an evidence-grounded agentic security operations harness on CloudflareThis post details the architecture of a new AI agent harness for security operations, moving from a single-agent approach to a multi-agent system. Key technical contributions include: 1. A 'recon first, inference second' approach where deterministic code collects a versioned evidence snapshot before AI agent analysis. 2. A lightweight triage model using Clef (running on Workers AI) to filter out high-volume noise. 3. A coordinator AI agent orchestrating four specialist AI agents (traffic analysis, customer context, global telemetry, threat intelligence) in parallel. 4. A synthesis AI agent that combines typed findings from specialists into an advisory report. 5. Use of Cloudflare's developer platform (Workers, Workflows, D1, R2, Durable Objects) for evidence collection, validation, orchestration, state persistence, and artifact storage. 6. Mechanisms for handling incomplete evidence and distinguishing between 'not checked', 'checked with no match', and 'checked with evidence supporting absence'.
  2. Cloudflare OS: your company’s agent workspace, managed for youThis post announces the waitlist for a fully managed Cloudflare OS deployment, simplifying the operational burden of the open-source agent workspace. It introduces new agent capabilities: direct integration with GitHub repositories for code exploration, bug fixing, feature addition, and pull request creation; enhanced Google Workspace integration via a Gatekeeper service for reading/researching Gmail, creating drafts, sending emails, and accessing Google Drive files; and expanded export formats for created content (Excel, CSV, PDF, Markdown, HTML).
  3. Detect and send production issues straight to your agentIntroduces 'Issues', a new built-in error monitoring system for Cloudflare Workers that automatically groups repeated exceptions, 5xx responses, and error logs into actionable issues. It enables sending these grouped issues, along with contextual data (stack traces, logs, traces, Worker version, user/account/session IDs via OpenTelemetry), to AI agents or other destinations via Automations. This streamlines the process of identifying, triaging, and fixing production issues by automating the handoff from observability to agent-driven remediation.
  4. How we’re rethinking work at Cloudflare with Cloudflare OSThis post details the internal development and rollout of Cloudflare OS, a platform for enabling safe and productive use of AI agents within Cloudflare. It outlines the guiding principles for AI adoption, including focusing on customer problems, empowering employees with 'superpowers', human ownership of AI output, context-aware AI, and strict permission controls. For engineers, it describes the Cloudflare Engineering Codex, an opinionated guide surfacing context across the SDLC, and AI agents for reviewing merge requests, technical designs, and incident reports, which have flagged nearly a quarter-million potential problems and blocked 16,000 merges. For non-engineers, it describes the evolution from a manual 'magic AI email bot' to a platform where identified 'jobs to be done' are automated through skill files and workflows, accessible via a web-based harness. Cloudflare OS V1 is a cloud-based harness running on Cloudflare's infrastructure, accessible via browser, authenticated through Cloudflare Zero Trust, and allows users to run skill files and workflows with outputs rendered in a panel. The platform emphasizes ephemeral cloud-based environments for security and auditability.
  5. How Cloudflare enforces engineering standards using AIThis post introduces the Cloudflare Codex, a governed set of engineering standards that agents can retrieve and apply at the point of work. It details the Codex organization, workflow (RFCs, domains, approval/enforcement states), and how statements are extracted into structured JSON for agent consumption. It describes three Codex consumers: an AI code reviewer that flags deviations and blocks merges based on enforced MUST requirements, a spec reviewer that evaluates technical designs against Codex requirements, and an incident report reviewer that checks for completeness and quality in postmortems. It also mentions alternative code review experiences like language-specific linters and a local CLI for the AI code reviewer.
  7. How we built a software factory to drive Astro’s GitHub issue count to zeroThis post details the evolution of an automated issue triage pipeline for the Astro repository, built on the Flue framework. It describes the development of an agent skill for reproducing, diagnosing, and fixing bugs, and its transformation into a state-machine-driven GitHub Action. The post highlights the benefits of this approach, including improved maintainer experience and the identification of architectural/documentation gaps. It also introduces the open-sourced triagebot-action and the Flue framework as adaptable tools for building similar agent automation pipelines.
  8. The Agent Development Lifecycle has arrived on CloudflareThis post introduces the Agent Development Lifecycle (ADLC) as a new paradigm for software development, replacing the traditional SDLC to accommodate the capabilities of AI agents. It details several new tools and initiatives from Cloudflare that support this vision: @cloudflare/ci for scalable CI/CD, local OpenTelemetry tracing for agents in development, Cloudflare Agents and Agent Traces for observing agent behavior, a system for enforcing engineering standards using AI, and a software factory for automating issue triage and bug fixing in open-source projects. The post emphasizes the need for programmatic, scalable, reproducible, real-time, atomic, permissioned, and self-improving platforms to enable 'software factories' driven by agents.
  9. Your agent needs a computer, not a container — introducing @cloudflare/computerIntroduces @cloudflare/computer, an open-source library providing an agent runtime that abstracts compute primitives (isolates, containers) into a unified 'computer' abstraction. It features a durable filesystem backed by SQLite and supports multiple execution runtimes, including isolates (via just-bash/dynamic workers) and containers (via Cloudflare Containers with FUSE). The package aims to enable more scalable agentic systems by providing a consistent environment for agents to interact with filesystems and tools, abstracting away the complexities of underlying compute infrastructure.
  10. Build your own vulnerability harnessThis post details the evolution of Cloudflare's AI-driven vulnerability discovery system, moving from a single-repository 'skill' to a fleet-wide, model-agnostic harness. Key contributions include: 1. Architectural shift to a model-agnostic harness, treating AI models as interchangeable components for continuous scanning and cross-testing. 2. Introduction of a two-stage operational framework: Vulnerability Discovery Harness (VDH) for proactive scanning and Vulnerability Validation System (VVS) for judgment and fixing, using different models for each stage to ensure adversarial validation. 3. Codification of the skill into a pipeline with distinct agents (Recon, Hunt, Validate, Gapfill, Dedup, Trace, Feedback, Report) managed by an orchestrator and backed by a database for persistence and state control. 4. Emphasis on strict context controls by keeping agent jobs hyper-focused to maintain context usage below 25% of the window. 5. Design for cross-repo dependency tracing to identify vulnerabilities across interconnected applications. 6. Strategies for eliminating false positives and coordinating end-to-end triage at scale.
  11. Bringing more agent harnesses and frameworks to Cloudflare, starting with FlueThis post introduces Flue, an open-source agent framework built on the Cloudflare Agents SDK. It details how Flue leverages the SDK's primitives for durable execution (runFiber, stash, onFiberRecovered), secure code execution via Dynamic Workers (@cloudflare/codemode), and a durable virtual filesystem (@cloudflare/shell). Flue's declarative approach and integration with existing platforms like Slack and GitHub are highlighted, along with its multi-cloud deployment capabilities, particularly its deployment as Durable Objects on Cloudflare for automatic scaling and isolated storage/compute.
  12. Announcing Claude Managed Agents on CloudflareThis post announces the integration of Claude Managed Agents with Cloudflare Sandboxes, enabling developers to run AI agent loops on Anthropic's platform while leveraging Cloudflare for code execution, security, and custom tool calls. It introduces Cloudflare's Developer Platform components like Sandboxes, Agents SDK, Browser Run, and Dynamic Workers as foundational elements for agent infrastructure. Key contributions include enhanced security via customizable proxies, detailed sandbox control and observability, lightweight sandbox options (V8 isolates via Dynamic Workers), private service connectivity using Cloudflare Mesh and Workers VPC, browser control and observability tools, email capabilities for agents, and custom tool execution. The integration provides a default deployment template and highlights the benefits of decoupling the agent's 'brain' (Anthropic) from its 'hands' (Cloudflare infrastructure).
  13. Building the agentic cloud: everything we launched during Agents Week 2026This post details numerous new product launches and enhancements related to AI agents and the 'agentic cloud'. Key announcements include Artifacts for Git-compatible versioned storage, Sandboxes for persistent and isolated agent environments, Outbound Workers for secure egress controls, Durable Object Facets for per-agent SQLite databases, and a rearchitected Workflows control plane for increased concurrency and creation rates. Security enhancements include Cloudflare Mesh for secure private networking, Managed OAuth for Access, and automated revocation/scoped permissions for non-human identities. The Agent Toolbox sees advancements with Project Think SDK, voice pipeline, Email Service beta, a unified inference layer, Unweight for LLM compression, Agent Memory, AI Search, and Browser Run. The Prototype to Production section introduces the unified 'cf' CLI, Agent Lee in-dashboard agent, Flagship feature flags, and beta for Cloudflare Registrar API. Finally, the Agentic Web section introduces the Agent Readiness score and Redirects for AI Training.
  14. Orchestrating AI Code Review at scaleThis post details the architecture of a CI-native orchestration system built around OpenCode, an open-source coding agent, which utilizes specialized AI agents for code review (security, performance, quality, documentation, release management, compliance) managed by a coordinator agent. The system processes merge requests, deduplicates findings, judges severity, and posts structured review comments, improving engineering resiliency and blocking merges for serious issues or vulnerabilities. It introduces a composable plugin architecture for VCS and AI provider integration, uses JSONL for structured logging from child processes, and implements a streaming pipeline for real-time processing of agent outputs. Specialized agents with tightly scoped prompts are used instead of a single large prompt.
  15. Redirects for AI Training enforces canonical contentIntroduces 'Redirects for AI Training', a new feature that leverages `cf.verified_bot_category` and HTML canonical tags to issue HTTP 301 redirects for verified AI training crawlers to up-to-date content. This feature aims to prevent AI models from being trained on deprecated documentation by enforcing canonical URLs. It also enhances Radar's AI Insights with Response Status Code Analysis for AI crawlers.
  16. Project Think: building the next generation of AI agents on CloudflareThis post introduces Project Think, a new set of primitives for building long-running AI agents on Cloudflare. Key contributions include: durable execution with fibers (crash recovery, checkpointing via `stash()` and `onFiberRecovered()`), sub-agents via Facets (isolated child Durable Objects with their own SQLite and typed RPC), persistent sessions with the Session API (tree-structured messages, forking, compaction, FTS5 search), and sandboxed code execution using Dynamic Workers and the codemode approach. The codemode approach replaces sequential tool calls with LLM-generated programs for increased efficiency and reduced token usage. Dynamic Workers provide a fast, memory-efficient sandbox for code execution with a capability model that grants explicit permissions.
  17. Browser Run: give your agents a browserThis post introduces Browser Run, a new product for AI agents to interact with the web. Key features include Live View, Human in the Loop, Chrome DevTools Protocol (CDP) Endpoint, MCP Client Support, WebMCP Support, Session Recordings, and increased concurrency limits. The CDP endpoint allows agents to control browsers directly, enabling existing CDP scripts and new agent frameworks to connect. WebMCP allows websites to expose tools for agents to discover and call, improving navigation reliability. The post also mentions existing Puppeteer, Playwright, Stagehand, and Quick Action endpoints, as well as a new /crawl endpoint.
  18. Introducing Agent Lee - a new interface to the Cloudflare stackThis post introduces Agent Lee, an in-dashboard AI assistant that interacts with the Cloudflare platform using natural language. It leverages Codemode to convert Cloudflare's MCP tool definitions into a TypeScript API, allowing the LLM to generate code for tool calls. A Durable Object acts as a credentialed proxy for sandboxed execution of generated code, classifying operations as read or write. Write operations are gated by an elicitation system requiring explicit user approval. Agent Lee utilizes generative UI components to dynamically render visualizations like charts and architecture maps alongside textual responses. The underlying primitives (Agents SDK, Workers AI, Durable Objects, MCP infrastructure) are available to customers.
  19. Deploy your own AI vibe coding platform — in one click!This post introduces VibeSDK, an open-source AI vibe coding platform that allows users to deploy their own end-to-end AI-powered development platforms. It integrates with LLM models for code generation, debugging, and iteration, leveraging Agents SDK. It utilizes Cloudflare Sandboxes for isolated development environments, Cloudflare Workers for Platforms for scalable deployment, and AI Gateway for observability, caching, and multi-model support. The platform also includes project templates stored in R2 and allows for one-click export to user accounts or GitHub repos. The post details the architecture, including sandbox creation, code generation, preview URL exposure, and deployment to Workers for Platforms. It also highlights the use of Google's Gemini models and the benefits of AI Gateway for routing requests across LLM providers, caching responses, and tracking costs.
  20. Automating threat analysis and response with CloudyThis post introduces the integration of Cloudy, Cloudflare's AI agent, with security analytics and a new conversational interface for faster root cause analysis and mitigation of traffic anomalies. Cloudy can now query Cloudflare's global threat intelligence data, providing contextual information about emerging threats, attacker activity, and indicators of compromise. This integration leverages the Agents SDK and Workers AI for rapid iteration and deployment, enhancing security analysis and response capabilities for users.
  21. Cloudy Summarizations of Email Detections: Beta AnnouncementThis post introduces the application of Cloudy, an AI-powered agent, to generate human-readable summaries of email detection rules. The challenge addressed is the communication gap between Cloudflare's rapid detection pipeline and SOC teams, who lack context on the specific factors triggering detections. The solution involves extending Cloudy's capabilities to translate complex ML model outputs and signal combinations into clear explanations. Key technical implementations include Retrieval-Augmented Generation (RAG) to ground LLM outputs in detection data and enhancing model context by clarifying internal model designations (e.g., 'Churchmouse' referring to sentiment detection models). The post details the 'hallucination problem' and the safeguards implemented to mitigate it, showcasing an example summary for the 'SPAM.ASNReputation.IPReputation_Scuttle.Anomalous_HC' rule. The beta program aims to rigorously test these outputs before wider release.
  22. Building agents with OpenAI and Cloudflare’s Agents SDKThis post details the integration of OpenAI's Agents SDK with Cloudflare's Agents SDK, which utilizes Durable Objects for persistent execution. It introduces the concept of a clear separation of concerns: OpenAI for cognition/planning and Cloudflare for execution/state. It showcases architectural patterns like multi-agent systems, human-in-the-loop workflows, and addressable agents, providing code examples for each. Key learnings include the importance of state persistence with Durable Objects, the role of routing in agent identity and lifecycle management, and the composability of agents. The post highlights the benefits of this combination for building intelligent, distributed agents.
  23. Building an AI Agent that puts humans in the loop with Knock and Cloudflare’s Agents SDKThis post details the integration of Knock's Agent Toolkit with Cloudflare's Agents SDK to build an AI agent that requires human approval for a virtual card issuing workflow. It showcases how to defer tool calls to Knock for human input, manage asynchronous approvals via webhooks, and leverage Durable Objects for stateful, globally addressable agent compute. The example demonstrates a practical application of AI agents in a human-in-the-loop scenario for financial services.