Living on the Frontier · Session 18 · In-person
Agents Clock In
June 27, 2026 · The Kannas Hotel, Chiang Mai
Last week the frontier went quiet.
This week it learned your name.
The agent stopped waiting to be asked.
It grew a memory. An identity. A seat at the table.
Notion @-mentions it like a coworker.
OpenAI says every department already feels it.
One inbox got shut down — its users had stopped opening it.
Underneath the launches, one shift:
the tool you used became the teammate you work with.
And the same week a face turned forgeable from a single photo —
because identity, it turns out, is the thing we’re all still learning to hold.
This week
Part I
The Headline Drops
The two frontier labs led the week: a model in three sizes, and an agent with a name.
OpenAI · Jun 26
GPT-5.6, in three sizes
OpenAI previewed GPT-5.6 Sol (next-gen frontier — stronger coding, science, cybersecurity), Terra (balanced, everyday), and Luna (fast, cheap, high-volume) — paired with their most advanced safety stack.
The naming is the story. One frontier brain, one workhorse, one cheap-and-fast — the same tiering we now choose between per task. The frontier isn't a single model anymore; it's a portfolio you route across.
"Be honest — what % of your work actually needs the frontier model versus the cheap-fast one?"
Anthropic · Jun 25
Claude Tag: a named, remembering agent
Anthropic shipped Claude Tag — a proactive, multiplayer agent with persistent memory and identity, built on top of Claude Code. Alongside it, a playbook for human–agent teams: the move from single-player AI to a team sport.
The load-bearing word is identity. A tool you invoke; a teammate you delegate to. Once the agent has a memory and a name, it stops being something you operate and becomes someone you manage — that's an org-design change wearing a product-launch costume.
"When your agent has a name and a memory, what's the first thing you stop doing yourself?"
Claude Tag — the reception
Karpathy: a 'third major redesign of LLM UI/UX'
This is a new paradigm for interacting with Claude that is significantly more 'inline' with all the other human activity org-wide.
— Andrej Karpathy · @karpathy
Part II
Teammates, Not Tools
The agent showed up where the team already works — @-mentioned, assigned, trusted to run.
Notion · Jun 24
Agents you @-mention like teammates
External Agents in Notion brings Claude and Cursor into the workspace your team already lives in. Assign them tasks from a shared board, @-mention them like colleagues, watch them run.
The agent shows up where the team already works, addressed the same way a human is. The board doesn't care whether the assignee has a pulse — and that's the whole point.
"If an agent and a person are both just @-mentions on a board, what actually changes about how you hand off work?"
OpenAI · Jun 25
'Every department' is going agentic
OpenAI says work across its entire company is being transformed by agents — Codex doing tasks that are more complex, longer-running, and increasingly cross-functional. They frame their own internal usage as an early look at what's coming for everyone.
Dogfooding as a roadmap. When the lab says we reorganized around this, that's the most honest benchmark there is — not an eval score, an org chart.
"Which of your weekly tasks is already 'longer-running and cross-functional' enough to hand to an agent?"
Notion · Jun 25
The inbox nobody opens anymore
Notion is winding down Notion Mail (Sep 22) — because more than half of its users now manage email without ever opening the inbox. Agents already run it, so Notion is "going all in" on that.
A product retiring itself because its users graduated past the UI. The endgame of a good agent isn't a better inbox — it's no inbox. Watch which of your apps are next to become "the thing the agent uses," not you.
"Which app do you still touch only because the agent can't fully run it yet?"
Part III
The Floor Moves Again
Computer use baked into the model, and silicon below a nanometer.
Google DeepMind · Jun 25
Computer use, built into the model
Gemini 3.5 Flash now ships native computer use — a built-in tool for agents that see and act across browser, mobile, and desktop. No bolt-on scaffolding; it's in the model.
Computer-use moving from a clever harness to a model primitive is the tell. When "click the screen" is a built-in, every app without an API just got one — the agent drives the GUI the way we do.
"What's one thing you do in a GUI every day that has no API — and would you let an agent click through it?"
IBM · Jun 25
Below one nanometer
IBM debuted the world's first sub-1nm chip — a 0.7nm / 7-ångström node with a new transistor architecture, claiming 70% greater energy efficiency.
Easy to scroll past a hardware number — but this is the floor under everything else. Cheaper, cooler compute is what makes "an agent per task, running all day" affordable. The model news rides on the silicon news.
"If inference got 70% cheaper, what would you run constantly that you currently run sparingly?"
Also shipped this week
Part IV
The Other Side of Identity
If software can have a self, so can your impersonator — and the org has to change shape.
Persona, via MTS · Jun 25
One photo is now enough to be you
Persona's CEO on the scariest shift in fraud: AI can now replicate a face from a single photo (it used to take hundreds). The game moved from fabricating synthetic identities to stealing real ones — so they're building credit-card-style alerts to warn the real you.
The dark mirror of "the agent got an identity": so did everyone's impersonator. The same week we gave software a self, a self became cheap to forge. Identity is quietly becoming the hard problem — for agents and humans both.
"If your face is forgeable from one photo, what do you trust instead — and does your agent know how to check?"
Steve Yegge · Jun 26
AI is not a compliance checkbox
Yegge on the brutal exec failure pattern: treating AI like SOC 2 — a checkbox to "get done in Q3" so the governance people are happy. His warning: your entire company is going to change shape, regardless of what leadership wants; the current hierarchy will be "nearly unrecognizable."
This is the thesis of the whole week, said plainly. Every card above — an agent with a name, every-department adoption, the inbox dying — is the company changing shape in real time. You don't "finish" AI in a quarter; you reorganize around it.
"If your org will be unrecognizable in two years — what shape do you actually want it to take?"
Also in the bookmarks
OpenAI · Jun 25
…and OpenAI gave houseplants a voice
The week's palate cleanser: openai/planttalk — an open-source build guide to give your houseplants a voice with ChatGPT. Yes, really.
Every serious week needs its joyful-stupid project. Also a tiny lesson: the gap between "a sensor + an LLM + a voice" and a shippable thing is now a weekend repo.
"What's the most useless-but-delightful thing you'd wire an agent into this week?"
Discussion
So — tool, or teammate?
Their entire company is going to change shape over the next few years, regardless of what they think, or want.
— Steve Yegge · this week
Follow the lab.
See you next Saturday.