Daily AI Zine
Friday, August 14, 2026
Issue No. 029 · Tokyo · full edition

✦ Today's Big Thing

GPT-5.6 makes agent work a routing problem

Today’s useful shift is not a single bigger model. It is smarter model selection, tighter agent lanes, and faster tests inside the tools Adrian already ships with.

5 min read · 7 sections

In Brief
  1. OpenAI’s GPT-5.6 builder guide points startups toward faster, more cost-efficient agents using smarter model selection and new Responses API capabilities.
  2. GitHub Copilot is adding Gemini 3.7 Flash, with early testing pointing to better web, app, and agentic development work.
  3. GitHub’s developer note reinforces the move from writing code to orchestrating the delivery system around code.
  4. StateFlow is a research signal for production and design: editable 3D world states are becoming a more serious previsualization target.

Today's Big Thing

The one thing that matters

Test it

GPT-5.6 shifts agent building toward lane routing

OpenAI’s latest builder guide is the most practical item today.

OpenAI’s builder guide for GPT-5.6 says startups are using the model family to build faster, more cost-efficient AI agents with smarter model selection and new Responses API capabilities. The important shift is operational: stop treating one model as the whole worker, and start routing jobs by difficulty, cost, and risk. For a daily builder, this means a customer-facing answer, a private draft, a repo edit, and a final review should not automatically use the same lane. Verdict: test one real workflow with explicit model lanes before changing the whole stack.

T1OpenAI News08/13

My AI Ecosystem

Your actual stack

Test it

Hot signal: GitHub Copilot adds Gemini 3.7 Flash

Gemini 3.7 Flash is rolling out in GitHub Copilot. GitHub says early testing showed improvements in web and app development and agentic development, meaning tasks where the assistant edits, checks, and follows through across more than one step. Treat this as a model-choice test inside existing repos, not a reason to change coding habits wholesale.

T1GitHub Changelog08/13
Use it

GitHub says developers are becoming orchestrators

GitHub’s new developer note argues that developers now own more of the delivery system around code, not only the code itself. Plain version: repo hygiene, prompts, reviews, tasks, releases, and agent supervision are becoming part of the job. This matches the daily reality of agent-assisted shipping, where the bottleneck is often direction and review.

T1GitHub Blog08/11
Watch it

StateFlow points previsualization toward editable 3D worlds

StateFlow is a research item for previsualization in film, games, architecture, and urban design. Its core idea is a 3D world state, meaning a saved scene structure with elements, positions, appearance, and time behavior that can be revised instead of regenerated from one prompt. Not ready as a tool pick, but relevant for design and production planning.

T1Hugging Face Daily Papers08/11

CoWork Corner

Claude CoWork, day to day

CoWork: turn tool-making into a repeatable room

Claude CoWork is the workspace Adrian uses for persistent projects, operational records, and AI-assisted workflows. No meaningful CoWork product change was found in today’s eligible Tier 1 items. The useful workflow to revisit is tool-making: Anthropic’s engineering index includes work on writing effective tools for agents with agents. In practice, keep one project room whose job is only to define repeated tasks, draft the small tool spec, name allowed files, and set a stop condition before Claude Code or Codex builds anything.

T1Anthropic Engineering Blog

GPT Desk

OpenAI, ChatGPT, Codex

Test it

Use GPT-5.6 as a lane router, not one big brain

What changed: OpenAI’s GPT-5.6 builder guide points to faster, more cost-efficient agents through smarter model selection and new Responses API capabilities. Why Adrian cares: Grey Group OS, Goodsense, and SET workflows all contain mixed-risk tasks, such as drafting, research, repo edits, and final client-facing language. What to do today: pick one workflow and define three lanes: cheap draft, strong reasoning, and final review. Run the same input through all three and save the cost, speed, and quality notes.

T1OpenAI News08/13
Use it

RingCentral turns ChatGPT Work and Codex into ops intelligence

What changed: OpenAI says RingCentral uses ChatGPT Work and Codex to accelerate AI product development and centralize operational intelligence across engineering and operations. Why Adrian cares: Grey Group OS should not only store project knowledge, it should turn that knowledge into decisions, blockers, and next actions. What to do today: create one Goodsense or SET decision log page, then ask ChatGPT to turn the last ten project notes into open questions, ownerless tasks, and reusable operating rules.

T1OpenAI News08/12

Small Money Systems

Small, repeatable, real

Test it

Small system: proposal upgrade sprint for Japan-facing work

System: build a GPT-5.6 proposal upgrade lane for short decks and bilingual outreach. Customer: small agencies, visiting producers, or Japan-facing brands that already have a rough brief. Offer: a cleaned proposal, sharper positioning, risk notes, and a follow-up email. Price: ¥30,000 per package. Existing assets: Goodsense, SET, Grey Group OS templates, past decks, and Japan-English business judgment. AI workflow: ChatGPT with GPT-5.6 model lanes handles rewrite, critique, and final polish, while Codex can turn the output into a repeatable form. First action: choose one old proposal and make the before-after sample today. Repeatability: every new brief runs through the same lane checklist. Effort: one hour. Expected value: small direct revenue and stronger lead conversion.

T1OpenAI News08/13
Test it

Small system: fast recap microsite for events and activations

System: create a repeatable recap microsite pack for small events, shoots, and brand activations. Customer: sponsors, venues, local brands, and collaborators around Street Attack Japan or Goodsense projects. Offer: a one-page recap with selected images, short copy, sponsor mentions, and next-step CTA. Price: ¥45,000 per recap. Existing assets: production planning habits, Netlify publishing, GitHub repos, and prior visual direction. AI workflow: GitHub Copilot’s new Gemini 3.7 Flash lane can be tested for the web and app build, while Claude Code or Codex keeps the template reusable. First action: make one dummy recap from an old event folder. Repeatability: each event swaps content into the same template. Effort: half day. Expected value: repeatable production revenue and sponsor follow-up material.

T1GitHub Changelog08/13

Build Next

Deployable now

Test it

Build a model-lane scorecard for one live workflow

What to build: a small scorecard page that records which model lane handled draft, reasoning, repo edit, and final review for one workflow. Why now: OpenAI’s GPT-5.6 guide puts smarter model selection and new Responses API capabilities at the center of agent building. Effort: one hour. Expected impact: lower waste and clearer quality review before scaling the workflow. Dependencies: an OpenAI API account, one existing repeatable task, and a place to save cost, speed, and output notes.

T1OpenAI News08/13
Test it

Build a Copilot Gemini branch test

What to build: a GitHub issue template that runs one web or app task through the usual coding lane and a Gemini 3.7 Flash Copilot lane, then records files changed, errors, and review notes. Why now: Gemini 3.7 Flash is rolling out in GitHub Copilot with reported improvements in web, app, and agentic development. Effort: one hour. Expected impact: better model choice for front-end work. Dependencies: GitHub Copilot access, one active repo, and a small task that can be safely repeated.

T1GitHub Changelog08/13

Try This Today

One action, right now

Pick one small repo or proposal task. Run it once through your normal lane, once through a GPT-5.6 lane, and once through Copilot’s Gemini 3.7 Flash if available. Save only three notes: time, quality, and what you would trust it with next.