Daily AI Zine · No. 017

Today's Big Thing

The one thing that matters

Use it

GitHub moves Copilot policy down to the user level

Enterprise AI administration is getting more granular.

GitHub put user-based model policy targeting for Enterprise customers into public preview. The practical change is simple: AI administrators with Copilot Business or Copilot Enterprise licenses can set a baseline and target model access by user, rather than treating the whole organization as one flat AI surface. That matters because Copilot is no longer one assistant with one risk profile. It is a routing layer for different models, agents, reviews, and workflows. Verdict: use it if an enterprise workspace already has mixed technical and non-technical users. For a smaller operation, copy the pattern manually: name who can use which model for which task, before the tool choice becomes habit.

T1GitHub Changelog07/31

My AI Ecosystem

Your actual stack

Use it

Copilot’s VS Code workbench keeps filling in agent basics

GitHub’s July Copilot releases for VS Code covered versions 1.127 through 1.131 and focused on agent work, reviewing changes, chat, and navigation. This is not a single flashy feature. It is the steady movement of coding agents into the normal editor loop. Verdict: use for active repos, especially where review and navigation friction still slows handoff.

T1GitHub Changelog07/30

Test it

OpenAI is making GPT-5.6 a cost-routing story

OpenAI says GPT-5.6 pricing is lower for Luna and Terra, positioning the family around price-performance rather than only frontier capability. The relevant shift is operational: routine agent work can be routed by cost, latency, and quality instead of defaulting to the strongest model. Verdict: test with repeatable tasks, not open-ended creative judgment.

T1OpenAI News07/30

Watch it

Hot signal: agent runaway guards are becoming a tool category

ITmedia reports that Perplexity open-sourced Numbat, tools intended to detect and prevent risky agent behavior before execution, with Claude Code and Codex named as integration targets. Anthropic’s own engineering index also puts product containment on the agent agenda. Verdict: watch the category, but test only in a sandbox until the tooling proves boring under real coding work.

T2ITmedia AI+07/31 T1Anthropic Engineering Blog

CoWork Corner

Claude CoWork, day to day

CoWork has no fresh product move, so tighten the tool map

Claude CoWork is the workspace Adrian uses for persistent projects, operational records, and AI-assisted workflows. No meaningful CoWork-specific product change showed up in today’s candidates. The useful revisit is still concrete: keep one short tool map per persistent project, listing which MCP-connected tools can read files, run code, or touch external systems. Anthropic’s engineering index continues to frame MCP code execution and context engineering as core agent topics, which makes this maintenance work worth doing before the next tool is added.

Tier 4 · quiet day, honest fallback

T1Anthropic Engineering Blog T1Anthropic Engineering Blog

GPT Desk

OpenAI, ChatGPT, Codex

Test it

Use GPT-5.6 pricing as a routing test for repeat work

OpenAI says GPT-5.6 pricing is lower for Luna and Terra. Adrian should care because Grey Group OS, Daily Ops, Goodsense proposals, and SET planning notes all contain repeatable language work where the best model is not always the strongest model. Do this today: create a 30-task sheet with three columns, brief cleanup, proposal polish, and implementation review, then run Luna or Terra against the current default and mark pass, edit, or fail.

T1OpenAI News07/30

Test it

Turn OpenAI’s voice-agent casework into a lead-revival script

OpenAI says Cars24 uses OpenAI-powered voice and chat agents for more than 1M monthly conversation minutes and to recover 12% of lost leads. Adrian should care because Goodsense, SET, Street Attack Japan, and Grey Group all have moments where warm inquiries go quiet after a proposal, shoot, event, or intro. Do this today: write one ChatGPT follow-up script for 20 dormant contacts, with a human approval step before anything is sent.

T1OpenAI News07/16

Small Money Systems

Small, repeatable, real

Test it

Small system: paid dormant-lead revival for production teams

System: build a human-approved follow-up agent that turns old inquiries, half-finished briefs, and post-event conversations into polite next-step messages. Customer: small agencies, production coordinators, venues, and Japan-entry teams that lose leads after first contact. Offer: a cleaned lead list, three message variants, and a seven-day follow-up tracker. Price: ¥35,000 for one batch. Existing assets: Grey Group contacts, Goodsense proposal language, SET production planning habits, Street Attack Japan activation follow-up patterns, and Grey Group OS templates. AI workflow: ChatGPT drafts and classifies, with manual approval before send. First action: pick 20 dormant contacts and paste only non-sensitive notes into a test sheet. Repeatability: every event, proposal, and shoot creates a new batch. Effort: one hour. Expected value: lead generation from work that already happened.

T1OpenAI News07/16 T1OpenAI News07/22

Use it

Small system: one-hour proposal upgrade pack

System: sell a fast AI-assisted proposal polish that improves structure, risk notes, budget language, and next-step clarity. Customer: founders, agencies, and production partners preparing a Japan-facing deck or client response. Offer: one revised proposal outline, one executive summary, and one follow-up email. Price: ¥50,000 per proposal. Existing assets: Goodsense deck instincts, Grey Group OS document patterns, SET planning language, and Adrian’s Japan-English business workflow. AI workflow: GPT-5.6 handles repeat drafting where lower pricing helps, while Codex can later turn the checklist into a small internal form. First action: choose one existing proposal and make the before-after template. Repeatability: each finished template becomes the starting point for the next client. Effort: one hour. Expected value: faster sales response and a paid service that also improves Adrian’s own proposals.

T1OpenAI News07/30

Build Next

Deployable now

Use it

Build a user-level Copilot policy register

What to build: a small repo file and GitHub issue template that records which users may use which Copilot models for drafting, code changes, review, and production-sensitive work. Why now: GitHub’s public preview for user-based model policy targeting makes this an actual admin pattern, not just an opinion. Effort: one hour. Expected impact: fewer accidental model choices and clearer review responsibility. Dependencies: a GitHub workspace, known user roles, and at least one active repo where Copilot is used.

T1GitHub Changelog07/31

Test it

Build a preflight guard for coding-agent runs

What to build: a lightweight preflight checklist that Claude Code or Codex must satisfy before running commands, touching credentials, creating packages, or pushing branches. Why now: reported Numbat work and Anthropic’s containment focus show that agent blast radius is becoming daily engineering hygiene. Effort: half day. Expected impact: fewer risky autonomous actions during repo work. Dependencies: an active coding-agent workflow, command logs, and a clear list of operations that always need human approval.

T2ITmedia AI+07/31 T1Anthropic Engineering Blog

Try This Today

One action, right now

Create one row with four fields: user, allowed model, allowed task, approval needed. Fill it for the next repo task before launching Copilot, Claude Code, or Codex. The point is not bureaucracy. It is preventing tool choice from becoming invisible.