Daily AI Zine · No. 023

Today's Big Thing

The one thing that matters

Test it

Hot signal: Cloudflare gives agents a browser built for them

Agent infrastructure is moving from demos to usable rails.

Cloudflare introduced Kitesurf, a stateless browser for AI agents that runs on Workers using V8 isolates, which are lightweight execution environments used to run many small tasks quickly. The point is practical: agents need browsers that are cheap to start, easy to scale, and designed for machine use rather than human browsing. This matters because web tasks are becoming the hands of agent workflows: research, form checks, page monitoring, and workflow testing. Verdict: test. Do not move production work onto it yet, but run one controlled browser task against a low-risk public page and compare speed, cost, and failure behavior against your current manual or script-based check.

T1Cloudflare Blog08/06

My AI Ecosystem

Your actual stack

Test it

WebMCP turns websites into agent-callable surfaces

Cloudflare launched a developer preview of WebMCP, described as a switch that makes a site usable by browser AI agents without new APIs or origin changes. MCP, or Model Context Protocol, is the connector layer that lets AI tools reach external systems in a structured way. Verdict: test. For any public site with repeated visitor tasks, this is worth a small sandbox rather than a production commitment.

T1Cloudflare Blog08/06

Test it

AI Search makes private knowledge retrieval less custom

Cloudflare AI Search lets developers point agents at their own files and websites to create a search layer, with a preview of pricing. The useful part is fewer primitives to stitch together before an agent can answer from owned material. Verdict: test. Use it first on non-sensitive public or prepared content, then compare answer quality against your current search or folder-based retrieval pattern.

T1Cloudflare Blog08/06

Use it

OpenAI’s scientific computing report reinforces the modernization queue

OpenAI published a field report on scientists using AI coding agents to modernize scientific computing work, including genomics. The direct takeaway is not the science domain. It is the pattern: old code, unclear dependencies, and useful but neglected workflows are now legitimate agent jobs when scoped tightly. Verdict: use. Build the backlog as small modernization tickets, not one broad refactor.

T1OpenAI News07/28

CoWork Corner

Claude CoWork, day to day

CoWork: no new move, so keep connectors narrow

Claude CoWork is the workspace Adrian uses for persistent projects, operational records, and AI-assisted workflows.

No meaningful CoWork-specific product change was found in today’s candidates. The useful workflow to revisit is still connector discipline: use Claude Desktop’s documented one-click MCP server installation path only for tools with a clear job, then keep a short note beside each project saying what the connector is allowed to touch. One vague connector can create more review work than it saves.

Tier 4 · quiet day, honest fallback

T1Anthropic Engineering Blog

GPT Desk

OpenAI, ChatGPT, Codex

Use it

Turn the Cars24 case into a missed-inquiry recovery test

OpenAI says Cars24 uses OpenAI-powered voice and chat agents for more than 1M monthly conversation minutes and to recover 12% of lost leads. Why Adrian should care: Goodsense, SET, and Street Attack Japan all benefit when old inquiries, quote requests, and half-finished conversations are reactivated without manual chasing. Do this today: export ten dormant inquiries, have ChatGPT draft three bilingual follow-up scripts, then manually send the best two before automating anything.

T1OpenAI News07/16

Test it

Give Codex one neglected workflow, not the whole repo

OpenAI’s scientific computing field report describes AI coding agents being used to modernize technical software. Why Adrian should care: Grey Group OS and Daily Ops likely contain small scripts, forms, templates, or glue steps that are valuable but too boring to clean up manually. Do this today: create one Codex ticket that only documents, tests, and improves a single repeated operations script or Netlify utility. No architecture work, no broad rewrite.

T1OpenAI News07/28

Small Money Systems

Small, repeatable, real

Test it

Small system: paid production search room

System: build a small searchable knowledge room for one project’s approved decks, briefs, schedules, and public reference links. Customer: a production company, agency producer, or Japan-market team that loses time finding the latest answer. Offer: a cleaned folder, an AI search layer, and a one-page usage guide. Price: ¥45,000 for setup. Existing assets: Grey Group OS structure, proposal habits, bilingual production planning, and reusable client handoff templates. AI workflow: use Cloudflare AI Search for the retrieval layer, with Claude Code or Codex preparing the file map and test questions. First action: pick one non-sensitive past project folder and write ten questions it should answer. Repeatability: each new client gets the same folder map and test script. Effort: half day. Expected value: paid setup fees plus better delivery speed on future projects.

T1Cloudflare Blog08/06

Use it

Small system: bilingual missed-inquiry rescue kit

System: create a lightweight follow-up kit for stale leads and unanswered inquiries. Customer: local service businesses, production vendors, and event teams with old messages sitting in email or forms. Offer: message triage, three bilingual reply scripts, a simple response log, and a one-week manual follow-up plan. Price: ¥30,000 for the first batch of up to 25 leads. Existing assets: Goodsense writing standards, SET production discipline, Street Attack Japan activation experience, and Adrian’s Japan-English workflow. AI workflow: ChatGPT drafts reply variants and Codex builds a simple tracking sheet or form. First action: paste five anonymized inquiries into ChatGPT and ask for reply categories. Repeatability: the same categories, scripts, and sheet work for every client. Effort: one hour. Expected value: small cash recovery work and warm lead generation.

T1OpenAI News07/16

Build Next

Deployable now

Test it

Build a browser-agent smoke test for one public workflow

What to build: a small test that asks an agent browser to open a public page, find one specific fact, and report whether a form or call-to-action is reachable. Why now: Kitesurf gives agents a browser designed for stateless, scalable web tasks rather than human browsing. Effort: one hour. Expected impact: faster QA on pages, campaigns, and public reference checks. Dependencies: access to Cloudflare’s agent-browser tooling, one safe public page, and a written pass or fail checklist.

T1Cloudflare Blog08/06

Use it

Write a WebMCP brief before touching a site

What to build: a one-page WebMCP brief that lists the site actions an agent may perform, such as reading a page, checking availability, or collecting a public brief, and the actions it may not perform. Why now: Cloudflare’s WebMCP developer preview makes agent access to websites a real implementation question. Effort: 15 minutes. Expected impact: fewer unsafe agent experiments and clearer client-facing site decisions. Dependencies: one candidate site, a basic task list, and someone responsible for approving the boundaries.

T1Cloudflare Blog08/06

Try This Today

One action, right now

Pick one public page you already maintain or check often. Write a three-step task: visit, find one fact, report one pass or fail result. Use it as the baseline for testing Kitesurf or WebMCP later, not as a production automation yet.