Daily AI Zine
Monday, July 27, 2026
Issue No. 013 · Tokyo · full edition

✦ Today's Big Thing

Contain the agent before you scale it

The useful shift today is not another assistant launch. It is the move toward bounded, reviewed, recoverable agent work.

4 min read · 7 sections

In Brief
  1. Anthropic is foregrounding agent containment as agents gain more access to products and tools.
  2. Google’s Gemini API managed agents now point toward background tasks and remote MCP connections.
  3. GitHub shows a pattern for turning merged product work into reviewed documentation pull requests.
  4. Runway’s Gen-4.5 is today’s creative hot signal, but it should be tested against real shot requirements.
  5. OpenAI’s coding-eval analysis is a reminder to judge Codex work by local tests, not leaderboard confidence.

Today's Big Thing

The one thing that matters

Use it

Agent containment becomes the practical story

Quiet day for broad platform launches, but the useful shift is clear: agent containment is becoming operating infrastructure, not a safety footnote. Anthropic is now foregrounding how it contains Claude across products, with the stated problem that more capable agents carry a larger potential blast radius. The practical takeaway is immediate: every autonomous workflow needs a capped action space, review points, and a rollback path before it is handed access to repos, files, deploys, or operational records. Treat new agent work as production operations. Small scoped permissions beat clever prompts.

T1Anthropic Engineering Blog

My AI Ecosystem

Your actual stack

Test it

Google managed agents add background and remote MCP patterns

Google says Gemini API managed agents are adding background tasks, remote MCP, and other capabilities for production-ready agents. The important part is not Gemini itself, but the direction: agents are moving from chat windows into long-running, tool-connected work. Test the pattern, not the platform.

T1Google AI Blog07/07
Use it

GitHub shows the docs-after-merge pattern

GitHub’s Agentic Workflows example turns merged product changes into subject-matter-reviewed documentation pull requests. That is a clean operating pattern: let the agent draft, but keep human review at the final gate. It fits any repo where releases move faster than documentation.

T1GitHub Blog07/08
Test it

Hot signal: Runway Gen-4.5 needs real shot tests

Runway is positioning Gen-4.5 as a new video model with stronger motion quality, prompt adherence, and visual fidelity. People are paying attention because those are the weak spots that decide whether generated video helps actual previsualization. Do not buy the claim abstractly. Test it against a real storyboard, reference frame, and motion note.

T1Runway Research

CoWork Corner

Claude CoWork, day to day

CoWork MCP boundaries are the useful revisit

Claude CoWork is the workspace Adrian uses for persistent projects, operational records, and AI-assisted workflows. No meaningful CoWork product change was found today. The useful workflow to revisit is MCP access design: keep one project-level note listing which tools can read, which can write, and which require manual approval. Anthropic’s engineering index includes code execution with MCP and agent containment, which makes the operating point plain: tool access is part of the workflow, not an afterthought.

T1Anthropic Engineering BlogT1Anthropic Engineering Blog

GPT Desk

OpenAI, ChatGPT, Codex

Use it

Use OpenAI’s eval-noise warning as Codex review discipline

OpenAI published an analysis raising concerns about reliability and accuracy in SWE-Bench Pro, a benchmark used to evaluate coding models. Adrian should care because Codex work inside Grey Group OS and Goodsense should be judged by local repo tests, not model reputation. Today’s action: create a Codex acceptance checklist with changed files, test command, expected behavior, rollback note, and one human review gate.

T1OpenAI News07/08
Watch it

Prepare a GPT-5.6 Sol sandbox, not a migration

OpenAI has previewed GPT-5.6 Sol as a next-generation model with stronger coding, science, and cybersecurity capabilities, paired with its most advanced safety stack. Adrian should care because it could become a better lane for Codex-heavy builds and API checks, but preview language is not an operating guarantee. Today’s action: write three fixed test prompts from Grey Group OS, SET, and Daily AI Zine work so the model can be judged quickly when available.

T1OpenAI News06/26

Small Money Systems

Small, repeatable, real

Use it

Sell a release-to-docs retainer

System: Build a release-to-docs service that turns shipped repo changes into reviewed documentation updates. Customer: small product teams, agencies, or internal ops teams that ship but do not document. Offer: monthly docs pull requests, changelog notes, and a short owner review queue. Price: 45,000 JPY per repo per month. Existing assets: Grey Group OS templates, GitHub habits, Netlify deployment knowledge, and Goodsense-style operating notes. AI workflow: Claude Code or Codex reads merged pull requests, drafts docs, and opens review-ready changes. First action: pick one existing repo and generate a sample docs PR today. Repeatability: every merge creates the next billable unit. Effort: half day. Expected value: a small monthly retainer and less unpaid explanation work after launch.

T1GitHub Blog07/08
Test it

Package an agent-containment audit

System: Build a fixed-scope AI agent containment audit. Customer: founders, production teams, and small companies starting to use ChatGPT, Codex, Claude Code, or MCP tools without clear permissions. Offer: a one-page risk map, access table, review gates, and rollback checklist. Price: 30,000 JPY for one workflow. Existing assets: Grey Group OS operating checklists, Adrian’s own Claude Code and Codex practice, and Daily Ops discipline. AI workflow: Claude Code drafts the audit template, Codex turns it into a reusable form, and ChatGPT rewrites the client-facing summary. First action: audit one Goodsense or Street Attack Japan workflow as the sample. Repeatability: each new client tool stack uses the same checklist. Effort: one hour. Expected value: paid advisory work and fewer avoidable agent mistakes.

T1Anthropic Engineering BlogT1OpenAI News07/08

Build Next

Deployable now

Use it

Build a local agent-containment gate

What to build: a small repo checklist that every Claude Code or Codex run must fill before changes are merged: files touched, tools used, tests run, rollback path, and approval status. Why now: Anthropic is foregrounding agent containment, and OpenAI is warning that coding evaluations can be noisy. Effort: one hour. Expected impact: fewer risky agent merges and faster review. Dependencies: an active GitHub repo, one coding-agent workflow, and a human reviewer willing to enforce the checklist.

T1Anthropic Engineering BlogT1OpenAI News07/08

Try This Today

One action, right now

Pick one active coding workflow. Write five lines: what the agent may read, what it may edit, what it must never touch, which command proves success, and who approves merge. Put it beside the repo before the next run.