--- title: Cloudflare Agents: tracing, local debugging, and a new... description: Cloudflare is launching a suite of tools for building and debugging AI agents on Workers. The Agents dashboard provides OpenTelemetry-compatible tracing with a... canonical: https://daily.dev/posts/cloudflare-agents-tracing-local-debugging-and-a-new-agent-runtime-8gcz1p1kb twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Cloudflare Agents: tracing, local debugging, and a new agent runtime | daily.dev og:description: Cloudflare is launching a suite of tools for building and debugging AI agents on Workers. The Agents dashboard provides OpenTelemetry-compatible tracing with a... og:url: https://daily.dev/posts/cloudflare-agents-tracing-local-debugging-and-a-new-agent-runtime-8gcz1p1kb og:image: https://api.daily.dev/og/posts/8Gcz1p1Kb.png og:image:alt: Cloudflare Agents: tracing, local debugging, and a new agent runtime og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # Cloudflare Agents: tracing, local debugging, and a new agent runtime **[Collections](https://daily.dev/sources/collections)** · 4 min read · 1 upvotes · 0 comments ## Summary Cloudflare is launching a suite of tools for building and debugging AI agents on Workers. The Agents dashboard provides OpenTelemetry-compatible tracing with a waterfall view of model calls, tool executions, and subagent delegations, with OTLP export support. Local debugging is now built into wrangler dev and vite dev — traces are captured at the workerd runtime level with no SDK changes, stored in SQLite-backed Durable Objects, and exposed via a REST API and browser UI. The third piece is @cloudflare/computer, an open-source agent runtime that gives each agent its own virtual machine with a durable SQLite-backed filesystem, transparently routing lightweight tasks to isolates and heavier tasks (git, shell, file ops) to Linux containers, targeting under 10% container usage for most workloads. ## Content Cloudflare ran an "Agents Week" in early August 2026, pushing out a wave of infrastructure, security, and tooling aimed at AI agents. Here's a consolidated look at what launched. ## @cloudflare/computer: a persistent runtime for agents The headline technical release is [@cloudflare/computer](https://github.com/cloudflare/computer), an open-source package that gives each AI agent its own virtual computer rather than a throwaway container. The core idea: most agent work is lightweight enough to run in Cloudflare's isolates (Durable Objects), and containers should only spin up when genuinely needed. Cloudflare's stated target is containers for less than 10% of agent workloads. What holds it together is a SQLite-backed virtual filesystem that works across all three execution backends: - **Worker shell** — a just-bash environment in a Dynamic Worker - **JavaScript Dynamic Worker** — ECMAScript modules - **Linux container** — full container sandbox for heavier tasks The filesystem splits files into 512 KiB SHA-256-hashed chunks. Before each container command, changes sync from the Durable Object into the container's FUSE mount; after the command, they sync back. Conflict resolution is last-writer-wins, deletes use tombstones, and pulls are batched with cursor checkpointing. Performance caveat worth knowing: the FUSE mount is competitive for metadata-heavy operations like git, but 30–40x slower than ext4 for large sequential reads. The workspace supports up to ~10 GB. The package includes an isomorphic-git client and Vercel AI SDK tool bindings. This is early preview — experiments and prototypes only for now. ## WriteGuard: policy and audit layer for MCP servers As agents start performing write actions at machine speed, Cloudflare built WriteGuard to sit in front of MCP servers and add controls without touching the underlying server code. Tools get assigned to one of four risk tiers: - **Read Only** - **Minimal Impact** - **Contained Write** - **Critical** — blocked before execution Beyond blocking, WriteGuard adds agent attribution to downstream applications so writes are traceable to both the human and the specific agent session. Audit events are logged asynchronously after scrubbing sensitive data, giving fleet-wide visibility across all 27 servers in Cloudflare's internal MCP portal (where this was originally built and tested). WriteGuard is entering private beta for external organizations. ## Cloudflare Agents: tracing and observability dashboard Cloudflare launched a unified dashboard for deploying and managing agents, with agent tracing as the first major feature. Tracing is OpenTelemetry-compatible and surfaces model calls, tool executions, token usage, and subagent delegation in a single waterfall view. Developers can replay full conversation sessions, inspect how agent operations connect to underlying infrastructure (KV, D1, Durable Objects), and export telemetry to any OTLP-compatible destination. For local development, `wrangler dev` and `vite dev` now automatically capture OpenTelemetry traces for local Worker invocations. When an agent session is detected, a Local Explorer API is surfaced automatically — agents can query structured traces, inspect binding state, and verify fixes without deploying. Traces are stored in a SQLite-backed Durable Object and exposed via a REST API with an OpenAPI schema. Supported frameworks at launch: Think, Flue, and AI SDK. Tracing is free during beta; pricing starts October 1, 2026 under existing Workers Observability tiers. ## Everything else from the week The week covered a lot of ground beyond the above: - **Agent Development Lifecycle (ADLC)** and programmable CI/CD - **Agent Wallets** for handling transactions - **Zero Trust extended to agents**, with AI activity attribution - **WebMCP and MCPv2**, an agent-optimized browser, and AEO (Answer Engine Optimization) - **Unified AI model routing** and an AI Search product - **$1M in new open-source funding** through Cloudflare Ambassadors and Community Engineers programs - **Radar AI research assistant** - Python/JS Worker interop and voice AI hosting The throughline across all of it: Cloudflare is betting that agents will be the primary users of internet infrastructure, and is building the plumbing accordingly — persistent compute, write controls, observability, and identity for agents rather than humans. ## Questions this post answers ### When does Cloudflare agent tracing stop being free and what pricing tier applies? Cloudflare agent tracing is free during beta, with pricing starting October 1, 2026 under existing Workers Observability tiers. The tracing captures every model call, tool execution, token count, and subagent delegation in a single waterfall view, and telemetry can be exported to any OTLP-compatible destination. _Teams planning agent infrastructure budgets track Cloudflare pricing changes like this on daily.dev._ ### What are the risk tiers in Cloudflare WriteGuard for MCP servers? WriteGuard defines four risk tiers for MCP tool actions: Read Only (no state changes), Minimal Impact (low-stakes writes), Contained Write (scoped mutations), and Critical (blocked before execution). The tiering works without modifying underlying MCP servers and logs every write with attribution to both the human user and the specific agent session. It was built internally for a portal running 27 MCP servers. _Engineers adding write tools to MCP servers are following WriteGuard's private beta rollout on daily.dev._ ### How does @cloudflare/computer handle state between isolates and containers for AI agents? @cloudflare/computer uses a shared SQLite-backed filesystem to bridge isolates and containers, so agents can hand off tasks between environments without losing state. Durable Objects serve as the primary agent harness, with Linux container sandboxes spun up only when needed — targeting less than 10% of work. Three backends are available: container projects (full Linux), isolate shells (just-bash), and isolate JavaScript (ECMAScript modules). _Developers experimenting with persistent agent compute environments are tracking @cloudflare/computer updates on daily.dev._ ## Similar posts on daily.dev - [Welcome to Agents Week](https://daily.dev/posts/welcome-to-agents-week-dvxe4ug2k) · Cloudflare · 1 upvotes · 0 comments --- Tags: [#ai-agents](https://daily.dev/tags/ai-agents), [#observability](https://daily.dev/tags/observability), [#cloudflare](https://daily.dev/tags/cloudflare), [#opentelemetry](https://daily.dev/tags/opentelemetry) [View this post on daily.dev](https://daily.dev/posts/cloudflare-agents-tracing-local-debugging-and-a-new-agent-runtime-8gcz1p1kb) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"Cloudflare Agents: tracing, local debugging, and a new agent runtime","url":"https://daily.dev/posts/cloudflare-agents-tracing-local-debugging-and-a-new-agent-runtime-8gcz1p1kb","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/cloudflare-agents-tracing-local-debugging-and-a-new-agent-runtime-8gcz1p1kb"},"datePublished":"2026-08-04T19:56:28.727Z","dateModified":"2026-08-10T18:38:35.123Z","description":"Cloudflare is launching a suite of tools for building and debugging AI agents on Workers. The Agents dashboard provides OpenTelemetry-compatible tracing with a...","isAccessibleForFree":true,"articleSection":"Collections","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Collections","logo":"https://media.daily.dev/image/upload/s--fk_6ycEi--/f_auto,q_auto/v1780996001/logos/collections?_a=BAMAMiWQ0","url":"https://daily.dev/sources/collections"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/cloudflare-agents-tracing-local-debugging-and-a-new-agent-runtime-8gcz1p1kb","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":1},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"ai-agents,observability,cloudflare,opentelemetry","timeRequired":"PT4M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Collections","item":"https://daily.dev/sources/collections"},{"@type":"ListItem","position":3,"name":"Cloudflare Agents: tracing, local debugging, and a new agent runtime"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/cloudflare-agents-tracing-local-debugging-and-a-new-agent-runtime-8gcz1p1kb#faq","mainEntity":[{"@type":"Question","name":"When does Cloudflare agent tracing stop being free and what pricing tier applies?","acceptedAnswer":{"@type":"Answer","text":"Cloudflare agent tracing is free during beta, with pricing starting October 1, 2026 under existing Workers Observability tiers. The tracing captures every model call, tool execution, token count, and subagent delegation in a single waterfall view, and telemetry can be exported to any OTLP-compatible destination. Teams planning agent infrastructure budgets track Cloudflare pricing changes like this on daily.dev."}},{"@type":"Question","name":"What are the risk tiers in Cloudflare WriteGuard for MCP servers?","acceptedAnswer":{"@type":"Answer","text":"WriteGuard defines four risk tiers for MCP tool actions: Read Only (no state changes), Minimal Impact (low-stakes writes), Contained Write (scoped mutations), and Critical (blocked before execution). The tiering works without modifying underlying MCP servers and logs every write with attribution to both the human user and the specific agent session. It was built internally for a portal running 27 MCP servers. Engineers adding write tools to MCP servers are following WriteGuard's private beta rollout on daily.dev."}},{"@type":"Question","name":"How does @cloudflare/computer handle state between isolates and containers for AI agents?","acceptedAnswer":{"@type":"Answer","text":"@cloudflare/computer uses a shared SQLite-backed filesystem to bridge isolates and containers, so agents can hand off tasks between environments without losing state. Durable Objects serve as the primary agent harness, with Linux container sandboxes spun up only when needed — targeting less than 10% of work. Three backends are available: container projects (full Linux), isolate shells (just-bash), and isolate JavaScript (ECMAScript modules). Developers experimenting with persistent agent compute environments are tracking @cloudflare/computer updates on daily.dev."}}]} ```