Skip to content

cloudflare-agent-stack-2026.md

Latest commit

 

History

History
93 lines (74 loc) · 5.43 KB

cloudflare-agent-stack-2026.md

File metadata and controls

93 lines (74 loc) · 5.43 KB
Raw

Cloudflare's AI / agent / "vibecoding" stack (2026) — what's relevant to PAGS

Status: research / reference, captured 2026-06-26. Forward-looking — relevant to the "open the platform for user-built agents" phase and to a cloud Coder placement. We are not adopting these yet; current priority is making the existing agents (Coder, apply) solid. Re-check versions before building.

Cloudflare shipped a full agent/code-execution stack in H1 2026. Several pieces directly bear on PAGS's architecture decisions.

1. Dynamic Workers (Dynamic Worker Loader API) — the "code on the fly, isolated" primitive

The most important one for a creator/AI-authored agent marketplace.

  • Hand the Workers runtime code as a string at runtime, get back an isolated V8 sandbox in single-digit milliseconds — no deploy pipeline. 
    const worker = env.LOADER.load({
  mainModule: "agent.js",
  modules: { "agent.js": agentCode },   // ← code as a string (from D1/R2)
  env: { ...scopedBindings },           // dynamic, per-tenant bindings
  globalOutbound: null,                 // block/inspect/rewrite outbound
});
  • 100× faster / cheaper than containers (few MB, ms startup) → an ephemeral sandbox per request/agent is viable at consumer scale; no concurrency cap.
  • Dynamic bindings: each tenant's code gets scoped tools via RPC/HTTP with secrets injected server-side (the code never sees your keys). "Every binding is heading for a dynamic counterpart — queues, DBs, object stores, AI, MCP."
  • Pricing: beta-waived; later $0.002 per unique Worker loaded/day + normal CPU/invocation. Constraint: JS/TS for best perf (Python/WASM slower).
  • Real platforms (e.g. Zite) use it to let users build apps via chat, each automation isolated in its own sandbox.

Implication for PAGS: this is the substrate for "agents live in the DB, code stored in the DB, written on the fly, run isolated, metered, paid by usage." It replaces "build your own sandbox" and is cheaper/simpler than per-Worker Workers-for-Platforms deploys for AI-generated/ephemeral agent code.

2. Cloudflare Sandbox SDK — incl. "Run Claude Code on a Sandbox"

A server-side sandbox with filesystem + shell/process + git. Published tutorial clones a GitHub repo into a sandbox and runs Claude Code server-side, returning a diff.

  • Partially obsoletes our earlier "CF can't host a shell → Coder must be local" assumption. A cloud Coder placement is theoretically possible per the Sandbox SDK — but it is NOT implemented (as of 2026-08-06 the shipped Coder is local-runner-only; no Sandbox integration exists in workers/api/src). Forward-looking only.
  • Tutorial shows discrete task runs (clone → task → diff/PR). Docs do not specify CPU/timeout/persistence/cost for long-lived interactive sessions → suited to autonomous task runs; unproven for persistent interactive sessions. Beta. See coder-cloud-vs-local-placement.md.

3. Code Mode — a better tool loop

Instead of many individual tool calls, the model gets one tool — codemode({ code }) — and writes TypeScript orchestrating all configured tools in a sandbox ("an entire API in 1,000 tokens"). @cloudflare/codemode rewritten (Feb 2026) to be runtime-agnostic.

Implication: candidate to replace our hand-rolled tool loop (lib/tools.ts, the max-3-round dedup loop) in the shared agent runtime — cheaper, more capable.

4. Cloudflare Agents SDK (v0.16.1, 2026-06-16)

First-party agent framework: Codemode runtime + connector model, durable CDP browser automation, Think orchestration, voice output-device selection. Plus Project Think ("next-gen AI agents on Cloudflare") and Dynamic Workflows ("durable execution that follows the tenant" = multi-tenant Workflows).

Implication: overlaps what we hand-built (AgentDO, the brain/hands, the JobApply/Coding workflows). Same situation as AI Gateway (see memory cf-platform-products-vs-custom-builds) — adopt-vs-keep evaluation rather than reinventing. Notably their browser automation is now "durable CDP" — compare to our browser-runner takeover machinery.

How this updates our plans

Our prior plan Update
Tier-2 creator code via Workers for Platforms deploys Lead with Dynamic Workers (Worker Loader) for on-the-fly/AI-generated code; WfP only for managed/custom-domain deploys.
Coder = local runner only Add a CF Sandbox cloud placement (headless, no laptop).
Hand-rolled tool loop Evaluate Code Mode.
Custom AgentDO / brain Evaluate aligning with the CF Agents SDK (like AI Gateway).

Sources