Skip to content
wojciech.io

Why this exists

Keeping up became a second job.

You know the feeling. Something important shipped, everyone is already talking about it, and you are finding out late. The models, agents, and tools under your stack change every week, and the one shift that matters is buried under launch threads and hot takes.

  • A dozen tools ship updates every week. Half of them change how you would build.
  • The shift that matters is buried under noise, reaction, and self-promotion.
  • By the time it reaches your feed, it is a week old and stripped of context.

AI Espresso is the fix. One operator reads the firehose every day and sends only what changes the work, in the time it takes to make an espresso.

What lands in your inbox

A real issue, not a mockup.

No editing, no cherry-picking. The real June 30, 2026 issue, exactly as it sent. It continues in your inbox.

AI Espresso · June 30, 2026 · hello@wojciech.io
AI Espresso
June 30, 2026 Wojciech

A daily operator briefing on AI models, agents, automation, and GTM workflows.

Operator briefing

Fast mode becomes architecture.

GitHub adds Opus 4.8 fast mode to Copilot as Anthropic retires it for 4.6, OpenAI folds Codex into ChatGPT Business seats, GitHub locks CI cache for untrusted triggers.

Routing

GitHub ships Opus 4.8 fast mode to Copilot.

It lands the same week Anthropic retires fast mode for 4.6. Model choice is turning into a routing decision you design, not a switch you flip.

Why it matters: Teams will pick models per task and per budget. The routing layer is the new architecture.

Source: github.blog

Seats

OpenAI folds Codex into ChatGPT Business seats.

Business workspaces can no longer add usage-billed Codex seats. One per-user price now covers ChatGPT and Codex, moving agent coding from metered experiment to standard seat.

Why it matters: Flat pricing makes agent coding a default, not a line item someone has to justify.

Source: help.openai.com

Supply chain

GitHub locks the Actions cache for untrusted triggers.

Forked pull requests get read-only cache, closing the path where an outside contribution poisons a cache that trusted CI later reuses.

Why it matters: Agentic CI runs more outside code. Cache isolation is now part of the threat model.

Source: github.blog

This is one issue. The thread continues every weekday morning.

Get it free →

Recent issues

Every issue, still here.

Nothing deleted, nothing quietly skipped. Judge the last 8 before you commit to the next one.

45 sent to date

  • Lifecycle Agents become an operational resource. AWS AgentCore moves agents into normal cloud workflows, Copilot Studio organizes them around knowledge and memory, Boomi ships an Agent Garden lifecycle API.
  • Runtime Production needs runtime, memory, a gate. Google frames the agent stack as services not one library, Tencent EdgeOne deploys agents like web apps, HubSpot wires G2 signals into Breeze Agents over MCP.
  • Access Top models arrive through gates. GPT-5.6 Sol signals frontier models dosed like critical infrastructure, GitHub Desktop 3.6 pulls in worktrees and Copilot, Vercel Agent Runs makes runs debuggable.
  • Control The workspace is set. Now the rules. Codex and Claude make the local agent a service you supervise from your phone, Vercel AI SDK 7 turns agents into a portable execution layer, Copilot for Jira closes ticket to PR.
  • Telemetry Agents get hands, now meters. Gemini 3.5 Flash gains native computer control, Heron shows what an agent actually does on the web, AI PR-spam turns into an operational problem.
  • Infra Agents become company infrastructure. Claude Tag turns Slack into shared delegation, Copilot App runs agent sessions on your own models, Mistral OCR 4 gives agents documents with layout not just text.
  • Trust An agent is not enough without a checker. OpenAI and Trail of Bits push fixes over more reports, GitHub brings enterprise agents to JetBrains with per-turn cost, Skybridge wants to be React for in-Claude apps.
  • Ops Control becomes the product. OpenAI gives admins one view of ChatGPT and Codex cost, GitHub closes the common pwn-request in Actions, Slackbot starts calling MCP tools from a channel.
Tomorrow's issue

Read tomorrow's issue before your first meeting.

One email a day, under three minutes, free. Currently 45 issues in. A full day of signal, one sip.

Subscribe free →