JA EN
shiichan

Agents SDK v0.14.0 lands: Agent Skills, messengers, and a lot more!

Cloudflare Changelog ★★★ View Markdown
#cloudflare#ai#serverless

Hey there, it's me, Shiichan! Today I found an update that makes building AI agents a whole lot more fun. Let me walk you through it.

Cloudflare Changelog developers.cloudflare.com

What was announced?

On Cloudflare's Changelog, Agents SDK v0.14.0 went live. It's a packed release: Agent Skills, messenger integration, scheduled tasks, Workflows integration, and hardened chat recovery all landed at once.

The main topics look like this:

  • Agent Skills (experimental)
  • Messengers (Telegram first)
  • Scheduled tasks
  • Think Workflows
  • Production hardening for durable chat recovery
  • MCP transport improvements

The story so far

Until now, giving an agent knowledge or skills usually meant cramming everything into the prompt, which kept ballooning. Chat integration was fiddly too — you handled webhooks yourself and nursed streams back to life when they stalled mid-turn.

What changes

v0.14.0 shifts a lot of that "you had to build it yourself" work over to the SDK.

With Agent Skills, you can load an on-demand catalog of instructions, so you call a skill only when you need it instead of fattening the prompt. Skills load from bundled directories, R2, or manifests, and expose activate_skill, read_skill_resource, and optionally run_skill_script tools. Sources that fail to load are skipped with a warning — a kind design.

Messengers connect a Think agent directly to chat platforms. Telegram is the first provider, and it handles webhook routing, conversation management, durable reply delivery, and streamed responses for you. See the Chat agents docs for the details.

Dive Deep

  • Scheduled tasks: write timezone-aware recurring prompts declaratively, like "every week on monday at 09:00" or "every day at 08:00 in Europe/London." It's backed by idempotent submissions that re-arm after each run.
  • Think Workflows: with ThinkWorkflow and step.prompt(), you can fold model reasoning steps into Cloudflare Workflows. It supports typed structured output via Zod schemas, long waits, and approval gates with configurable timeouts.
  • Chat recovery hardening: turns now survive deploys and evictions without losing completed work. useAgentChat gains an isRecovering flag so your UI can show a recovery state, stalled streams recover via chatStreamStallTimeoutMs, and sub-agents re-attach to results instead of re-running when the parent recovers.
  • MCP got better too: resumable SSE streams with Last-Event-ID reconnection, readable server IDs, and improved request correlation across HTTP and RPC.

Installing it is one line:

npm i agents@latest @cloudflare/think@latest @cloudflare/ai-chat@latest

For the finer details, the Agents API reference has you covered.

Wrap-up

  • Agents SDK v0.14.0 adds Agent Skills, messengers, scheduled tasks, and Workflows integration
  • Chat recovery is hardened for production, so turns survive deploys and evictions
  • MCP transport improved around reconnection and IDs
  • Install is a single npm i agents@latest @cloudflare/think@latest @cloudflare/ai-chat@latest

If you're building agents or chatbots on Cloudflare Workers, this update is right up your alley!

Related posts