Daily AI Zine
Sunday, September 6, 2026
Issue No. 052 · Tokyo · full edition

✦ Today's Big Thing

Route the expensive agents before they route the work

The useful shift today is not one more model name. It is the need to assign high-end agents to the few jobs where their cost, safety posture, and autonomy actually pay back.

6 min read · 7 sections

In Brief
  1. Anthropic’s Fable 5.1 and Mythos 5.1 put coding and knowledge-work routing back on the operating agenda.
  2. OpenAI’s Astra safety notes make auditability part of model selection, not an afterthought.
  3. Cloudflare’s BotBase update signals a more formal internet for agents, with behavior declarations and submission tracking.
  4. Codex case studies keep pointing toward one practical pattern: make business teams submit build requests in a shape agents can execute.

Today's Big Thing

The one thing that matters

Test it

High-end agent work is entering the routing phase

Anthropic has announced Claude Fable 5.1 and Claude Mythos 5.1 as advanced models for coding and knowledge work. The practical change is that agent work now needs routing, meaning each task gets assigned to the cheapest capable model rather than the newest model by default. A Japan report says Fable 5.1 may reduce some costs by up to about 45% through lower cache-read pricing, which matters for long sessions where the same project context is reused. Treat this as a test, not a blanket migration: define three lanes, quick edits, trusted production changes, and deep research, then measure which lane actually earns the premium model.

T1Anthropic NewsroomT2ITmedia AI+09/05

My AI Ecosystem

Your actual stack

Use it

Codex keeps moving from developer tool to business intake layer

OpenAI’s loveholidays case says Codex helped make software development accessible across the business, so more employees could turn ideas into products faster. The useful pattern is not “everyone codes.” It is structured intake: business people describe a workflow, Codex turns it into a draft, and a technical owner reviews the result before it ships.

T1OpenAI News08/26
Use it

AI-native workflow cases point to operating capability, not isolated prompts

OpenAI describes Basis, Clay, and Exa Labs using agents across onboarding, account management, and developer integrations. The common thread is repeatable workflow ownership: an agent does a defined slice of work, inside a business process, with humans still accountable for results. That is more useful than another prompt library.

T1OpenAI News09/01
Watch it

Hot signal: agent operators are being asked to declare behavior

Cloudflare’s BotBase update gives bot and agent operators a dashboard path for submissions, status tracking, editing, and declaring how their bots use content. The signal is clear: as agents browse and act online, access may increasingly depend on declared behavior and ongoing trust, not just technical evasion or user-agent strings.

T1Cloudflare Blog08/28T1Cloudflare Blog08/07
Watch it

Astra’s safety threshold makes model choice an approval issue

OpenAI says Astra is its first model to meet the Critical cybersecurity capability threshold under its Preparedness Framework, with stronger safeguards for release. In plain terms, this is a more capable model in a riskier class. For production work, access policy, logging, and task boundaries matter as much as benchmark performance.

T1OpenAI News09/01T2Perplexity Research09/03

CoWork Corner

Claude CoWork, day to day

CoWork has no fresh product signal today

Claude CoWork is the workspace Adrian uses for persistent projects, operational records, and AI-assisted workflows. No meaningful CoWork product change showed up in today’s candidates. The useful revisit is tool boundary discipline: Claude Desktop has had one-click MCP server installation, which makes adding tools easier, but easier installation also makes room sprawl easier. Keep each workspace tied to a short tool manifest: what the room can access, what it must not touch, and which human owns the handoff.

Tier 4 · quiet day, honest fallback

T1Anthropic Engineering Blog

GPT Desk

OpenAI, ChatGPT, Codex

Test it

Put Astra in an audit lane before production work

OpenAI says Astra has crossed a Critical cybersecurity capability threshold and is being released with stronger safeguards. Adrian should care because Grey Group OS, Codex, and client-facing proposal work can all create high-trust artifacts where invisible agent behavior is a liability. Today, create one Astra audit lane: allow research summaries and code suggestions, require saved prompts and outputs, and block autonomous changes to production repositories until the lane has passed one controlled task.

T1OpenAI News09/01T2Perplexity Research09/03
Use it

Give Codex a business-request form, not a vague build brief

OpenAI’s loveholidays case frames Codex as a way for non-developers to turn ideas into products faster. Adrian should care because Goodsense, SET, Street Attack Japan, and Grey Group OS all generate small workflow ideas that die when they become loose Slack-style asks. Today, make a Codex intake form with five fields: goal, user, current manual steps, success check, and repo or file location.

T1OpenAI News08/26
Use it

Use ChatGPT Work for deadline asset production

OpenAI says Stampli used Codex and ChatGPT Work to compress launch production, cutting launch hours by 68% under a fixed deadline. Adrian should care because decks, production briefs, recaps, and sponsor materials often have immovable delivery windows. Today, pick one current proposal or recap and run a two-pass workflow: ChatGPT Work drafts the asset set, Codex or a repo agent formats the reusable template.

T1OpenAI News08/20

Small Money Systems

Small, repeatable, real

Use it

Small system: Codex intake clinic

System is a one-hour Codex intake clinic that turns a messy business workflow into a build-ready task card. Customer is a small founder, agency producer, or Japan-market operator with repeated manual admin. Offer is one mapped workflow, one Codex-ready brief, and one review checklist. Price is ¥55,000. Existing assets are Grey Group OS, Daily Ops, Goodsense proposal structure, and Adrian’s Codex practice. AI workflow uses ChatGPT to interview the customer and Codex to draft the implementation task. First action is to make one intake template today. Repeatability comes from reusing the same template for every client. Effort is one hour. Expected value is paid diagnosis now and a clean path to follow-on build work.

T1OpenAI News08/26
Use it

Small system: launch-deadline rescue pack

System is a fixed-scope rescue pack for launch pages, decks, captions, recaps, and email copy when a deadline is close. Customer is a startup, brand team, or production partner that has content fragments but no finished asset set. Offer is a cleaned asset map, revised copy, and export-ready delivery checklist. Price is ¥88,000. Existing assets are SET production planning, Goodsense deck patterns, and Grey Group OS templates. AI workflow uses ChatGPT Work for draft production and Codex for template formatting. First action is to choose one past launch and turn it into the sample before-and-after. Repeatability comes from the same intake, asset map, and checklist. Effort is half day. Expected value is immediate service revenue and stronger upsell into monthly production support.

T1OpenAI News08/20
Test it

Small system: agent behavior readiness scan

System is a lightweight scan for companies planning to run bots, scrapers, or AI agents against public web properties. Customer is a local business, media operator, or SaaS team that wants agents to act online without getting blocked or creating reputational risk. Offer is a behavior declaration draft, risk notes, and a recommended access policy. Price is ¥66,000. Existing assets are Grey Group OS operating checklists and Adrian’s AI workflow audit capability. AI workflow uses Claude or ChatGPT to interview the operator, then Codex to generate the policy template. First action is to draft a one-page BotBase-style questionnaire. Repeatability comes from reusing the scan for each operator. Effort is one hour. Expected value is paid advisory revenue and leads for deeper automation work.

T1Cloudflare Blog08/28T1Cloudflare Blog08/07

Build Next

Deployable now

Use it

Build a Codex intake-to-task card

What to build is a simple form that converts a business request into a Codex-ready task card with goal, files, constraints, tests, and reviewer. Why now is OpenAI’s latest Codex case showing business users becoming builders when intake is structured. Effort is one hour. Expected impact is fewer vague build requests and faster review. Dependencies are an existing repository, a Codex account, and one human reviewer who can accept or reject the output.

T1OpenAI News08/26
Test it

Build an agent behavior declaration template

What to build is a reusable declaration template for any agent that browses, scrapes, posts, or calls external services. Why now is Cloudflare’s BotBase update, which formalizes submission tracking and behavior declarations for bots and agents. Effort is one hour. Expected impact is lower operational risk before deploying agents online. Dependencies are a real agent workflow to document and agreement on what content it may access or use.

T1Cloudflare Blog08/28

Try This Today

One action, right now

Take one messy automation idea and fill five fields only: goal, user, current manual steps, success check, and files touched. If the task still feels vague after that, it is not ready for an agent.