Written from 15 named sources GROUPED RELEASE BRIEFING (a) Compute & Durability 1. Workflows V2 (Deterministic Execution) Description: A major architectural overhaul of Cloudflare’s orchestration engine, moving to a deterministic, replayable step-based model with massive concurrency limits. Status & Date: GA — May 15, 2026 [13]. Concrete Primitive: WorkflowStep execution context; limits increased to 50,000 concurrent instances and 2 million queued instances [2], [13]. Why it matters for our stack: This is the direct, native solution to our 14-minute DO alarm constraint. Instead of fighting wall-clock execution limits in a DO alarm loop, Workflows V2 allows our agentic pipeline to sleep, wait for external HTTP fetches, or pause for OpenRouter LLM responses over days or weeks without timing out. Risk or Gotcha: Migration requires a full code restructure. You cannot simply wrap existing long-running DO code; tasks must be broken into idempotent steps [13]. Citation: [13] 2. Durable Object (DO) Facets Description: Enables "Supervisor" Durable Objects to instantiate "Facet" child objects with isolated SQLite databases on-the-fly without pre-provisioning namespaces. Status & Date: GA — April 2026 [6]. Concrete Primitive: this.ctx.facets.get(dynamicClassName) [6]. Why it matters for our stack: This solves multi-tenant data isolation for our agentic pipeline. We can spin up a dedicated, isolated stateful environment (with its own 10GB SQLite DB) for a specific agent task or tenant dynamically, keeping agent memory strictly partitioned. Risk or Gotcha: Facets are permanently tied to their supervisor; if the supervisor is deleted or its namespace compromised, access to the child facets is broken [6]. Citation: [6] (b) Observability 3. Browser Run: Live View & Session Replay Description: Real-time visual debugging and historical session recording for headless browser automation. Status & Date: GA — April 15, 2026 [4], [10]. Concrete Primitive: Dashboard "Runs" tab or the live.browser.run URL for active sessions [3], [4]. Why it matters for our stack: This closes a massive observability gap in our agentic pipeline. When an LLM-driven Browser Run task fails to click a DOM element, Sentry only gives us the stack trace. Live View allows us to visually replay the session to see exactly how the DOM mutated before the failure. Risk or Gotcha: Session recordings capture the visual state of the page, which may contain PII. Ensure agents are not navigating to sensitive user data if global recording is enabled [4]. Citation: [4], [10] (c) AI / Agents 4. Browser Run (formerly Browser Rendering) Rebrand & Scale Description: A pivot of the headless browser service to specifically target AI Agents, moving to Cloudflare Containers for faster cold starts and higher throughput. Status & Date: GA — April 15, 2026 [4], [8]. Concrete Primitive: env.BROWSER binding; concurrency increased to 120 sessions per account [5]. Why it matters for our stack: Our parallel agentic pipeline is heavily constrained by browser throughput. The 4x increase in concurrency (from 30 to 120) allows us to scale our headless automation tasks without hitting immediate 429s. Risk or Gotcha: Pricing scales aggressively. Beyond the initial 10 concurrent sessions (averaged monthly), additional concurrent browsers cost $2.00 each [7]. Citation: [4], [5], [7] 5. WebMCP Support in Browser Run Description: Integration of the Model Context Protocol (MCP) into Browser Run, allowing target websites to expose structured tools directly to agents. Status & Date: Beta — April 2026 [8]. Concrete Primitive: Native MCP tool discovery within the Browser Run session context [8]. Why it matters for our stack: It allows our agentic pipeline to interact with supported web pages via structured APIs rather than fragile DOM scraping, vastly improving the reliability of the headless browser phase. Risk or Gotcha: Utility is entirely dependent on target websites implementing MCP; it provides no benefit for scraping legacy or non-compliant web properties [8]. Citation: [8] (d) Storage & Data 6. SQLite-backed DO Billing & 10GB Limit Description: Formalization of the SQLite storage backend for Durable Objects, including a massive capacity increase and the activation of live billing. Status & Date: Billing Live — January 7, 2026 [1]. Concrete Primitive: ctx.storage.sql; storage limit increased to 10GB per object [1], [9]. Why it matters for our stack: Our stack relies heavily on SQLite storage for agent memory. The 10GB limit ensures we can maintain massive local state and vector caches directly inside the DO without externalizing to D1. Risk or Gotcha: Billing is now active. High-churn SQLite databases (frequent writes/deletes by agents) will incur significant costs based on "Storage API" operations [1], [11]. Citation: [1], [9] 7. Vectorize: Metadata Filtering & GA Description: Cloudflare’s vector database for RAG, now supporting advanced metadata filtering. Status & Date: GA — Late 2025 [13]. Concrete Primitive: env.VECTORIZE.query(vector, { filter: { tenant_id: "..." } }). Why it matters for our stack: Essential for the memory retrieval phase of our agentic pipeline. We can store embeddings from OpenRouter models and use metadata filters to guarantee agents only retrieve context authorized for their specific tenant. Risk or Gotcha: Strict dimension limits apply; we must ensure the embedding models we route through OpenRouter match the pre-configured Vectorize index dimensions [13]. Citation: [13] 8. Durable Objects: Point-in-Time Recovery (PITR) Description: The ability to restore SQLite-backed DOs to any exact microsecond state within the last 30 days. Status & Date: GA — Early 2026 [14]. Concrete Primitive: ctx.storage.restore(timestamp) or via the Cloudflare API [14]. Why it matters for our stack: Provides a critical safety net for our stateful agents. If an LLM hallucinates and corrupts its local SQLite database via sandboxed code execution, we can programmatically roll back the DO to the state right before the bad execution. Risk or Gotcha: This feature is exclusive to SQLite-backed DOs; legacy KV-backed DOs cannot use PITR [14]. Citation: [14] (e) Edge / Networking 9. AI Gateway: Request Tracing & Cost Controls Description: An edge proxy layer for LLM providers that provides caching, rate limiting, and unified logging. Status & Date: GA — Late 2025 [8]. Concrete Primitive: https://gateway.ai.cloudflare.com/v1/{account_id}/{gateway_name}/openrouter endpoint prefix. Why it matters for our stack: Since we use OpenRouter exclusively, routing those calls through AI Gateway gives us Sentry-adjacent observability for LLM latency, token usage, and costs, plus edge caching for repeated agent prompts. Risk or Gotcha: Adds a slight latency penalty to every LLM call and requires updating our OpenRouter base URL across all agent configurations [8]. Citation: [8] (f) Developer Experience 10. Unified cf CLI & Local Tunnels Description: A new unified CLI replacing parts of Wrangler, featuring improved local-to-remote tunneling for development. Status & Date: GA — May 2026 [3]. Concrete Primitive: cf dev command with automatic tunnel creation. Why it matters for our stack: Simplifies testing our SSE streaming and DO interactions locally. The dev server can securely tunnel into production DO namespaces, allowing us to test agentic pipelines against real state without deploying. Risk or Gotcha: The legacy Wrangler --remote flag was officially deprecated on May 18, 2026. Any CI/CD pipelines or local package.json scripts still using it are now broken [3]. Citation: [3] QUICK-REFERENCE MATRIX TABLE Release Name Status Date Key Primitive Relevant Stack Layer Risk Level Workflows V2 GA 2026-05-15 WorkflowStep Compute (Execution) High (Refactor) DO Facets GA 2026-04-01 ctx.facets.get() Durability (State) Low Browser Run Live View GA 2026-04-15 live.browser.run Observability Low Browser Run Scale GA 2026-04-15 120 Concurrency AI / Agents (Automation) Med (Cost) WebMCP Support Beta 2026-04-01 Native MCP discovery AI / Agents (Automation) Med (Adoption) SQLite DO Billing Live 2026-01-07 ctx.storage.sql Storage & Data Med (Cost) Vectorize Filtering GA 2025-11-01 env.VECTORIZE.query() Storage & Data Low DO PITR GA 2026-02-10 storage.restore() Storage & Data Low AI Gateway GA 2025-11-01 gateway.ai.cloudflare.com Edge / Networking Low Unified cf CLI GA 2026-05-01 cf dev Developer Experience Med (Breaking) TOP 5 TO ACTUALLY ADOPT 1. Workflows V2 Ranking Rationale: #1 because it directly eliminates our hardest platform constraint. The Tradeoff (DO Alarms vs. Workflows V2): Staying on DO alarms gives us the convenience of keeping agent context in memory between alarm ticks (assuming V8 doesn't evict the isolate), but we constantly battle the 14-minute wall-clock death. Migrating to Workflows V2 gives us infinite execution time and native retries, but forces us to serialize state between WorkflowStep boundaries. Given our reliance on OpenRouter and slow DOM interactions, Workflows V2 is the correct architectural choice. Smallest Concrete First Step: Install the updated types and define the workflow binding in wrangler.toml. # wrangler.toml [[workflows]] name = "agent-executor" binding = "AGENT_WORKFLOW" class_name = "AgentWorkflow" // index.ts stub import { WorkflowEntrypoint, WorkflowStep, WorkflowEvent } from 'cloudflare:workers'; export class AgentWorkflow extends WorkflowEntrypoint<Env, any> { async run(event: WorkflowEvent<any>, step: WorkflowStep) { const llmResponse = await step.do('call-openrouter', async () => { // Fetch logic here }); } } 2. Browser Run (Native CDP & Concurrency) Ranking Rationale: #2 because it immediately unblocks our parallel agentic pipeline throughput (jumping from 30 to 120 concurrent sessions) with minimal code changes. Smallest Concrete First Step: Update the Playwright provider package to ensure we are using the native CDP protocol rather than the legacy chunked protocol. npm install @cloudflare/playwright@latest 3. Durable Object Facets Ranking Rationale: #3 because it drastically simplifies our SQLite schema. Instead of writing complex WHERE tenant_id = ? logic for every agent memory query, we isolate state at the infrastructure level. Smallest Concrete First Step: Update your Supervisor DO to instantiate a facet for a new agent task. // Inside Supervisor DO async startAgentTask(taskId: string) { // Creates or retrieves an isolated DO with its own 10GB SQLite DB const taskFacet = await this.ctx.facets.get(AgentTask_${taskId}); return taskFacet.fetch("http://internal/start"); } 4. Browser Run Live View Ranking Rationale: #4 because it provides immediate ROI on observability without requiring code changes. It supplements Sentry by showing us why an agent failed visually. Smallest Concrete First Step: No code required. Instruct the team to navigate to the Cloudflare Dashboard -> Workers & Pages -> Browser Run -> "Runs" tab during active agent testing to view the live CDP stream. 5. Durable Object PITR Ranking Rationale: #5 because it provides a zero-config safety net for our stateful agents. Agentic pipelines are non-deterministic; if sandboxed code executes a destructive SQL query, we can undo it. Smallest Concrete First Step: Add the restoration command to our internal runbooks for on-call engineers. # Runbook snippet for DO rollback npx wrangler durable-objects storage restore <DO_ID> --timestamp="2026-05-18T14:30:00Z" Sources [1] Durable Objects, Workers - Billing for SQLite Storage - Changelog — https://community.cloudflare.com/t/durable-objects-workers-billing-for-sqlite-storage/869004 [2] Increased Workflows instance and concurrency limits · Changelog — https://developers.cloudflare.com/changelog/post/2025-10-28-raising-limits/ [3] Changelog · Cloudflare Browser Run docs — https://developers.cloudflare.com/browser-run/changelog/ [4] Browser Rendering is now Browser Run · Changelog — https://developers.cloudflare.com/changelog/post/2026-04-15-br-rename/ [5] Browser Run: now running on Cloudflare Containers, it’s faster and more scalable — https://blog.cloudflare.com/browser-run-containers/ [6] Durable Object Facets · Cloudflare Dynamic Workers docs — https://developers.cloudflare.com/dynamic-workers/usage/durable-object-facets/ [7] Pricing · Cloudflare Browser Run docs — https://developers.cloudflare.com/browser-run/pricing/ [8] Cloudflare Rebrands Browser Rendering to Browser Run for AI Agents StartupHub.ai — https://www.startuphub.ai/ai-news/technology/2026/cloudflare-rebrands-browser-rendering-to-browser-run-for-ai-agents [9] Release notes · Cloudflare Durable Objects docs — https://developers.cloudflare.com/durable-objects/release-notes/ [10] Browser Run: give your agents a browser — https://x.com/Cloudflare/status/2044415638165184973 [11] Pricing · Cloudflare Durable Objects docs — https://developers.cloudflare.com/durable-objects/platform/pricing/ [13] Cloudflare Introduces Workflows V2 with Deterministic... — https://app.daily.dev/posts/cloudflare-introduces-workflows-v2-with-deterministic-execution-and-50k-concurrent-workflows-kfrnp4drc [14] Access Durable Objects Storage · Cloudflare Durable Objects docs — https://developers.cloudflare.com/durable-objects/best-practices/access-durable-objects-storage/ Cloudflare Workflows scales with higher concurrency limits - LinkedIn — https://www.linkedin.com/posts/cloudflare_rearchitecting-the-workflows-control-plane-activity-7450203843990233088-gwN_ Concurrent Workflow instances limits increased. · Changelog — https://developers.cloudflare.com/changelog/post/2025-02-25-workflows-concurrency-increased/