Events
Talks from the frontier. Decks and notes from Living on the Frontier and Agents in the Wild — free to read, free to present.
The Discussion Board
Living on the Frontier, Session 23 — a format experiment. No news recap this week: the audience asked for technical discussion, so the community submitted the agenda all week in the group chat. Fifteen topics on the board, the room votes, we talk. Spec-driven development and whether code is still the source of truth, models that delete code versus models that route around it, the harness debates, sandbox escapes, and the economics of commodified intelligence.
Open Source Accelerates
Living on the Frontier, Session 22 — a one-week window in which the frontier got cheaper and less contained. DeepSeek opened V4-Flash's API in public beta with agent scores past its own Pro preview. Kimi K3 published 2.8 trillion parameters on a Monday, and by that same evening it was live on a gateway, an editor, an agent tool and a serving library — every one of them leading with 'US-hosted, zero data retention.' OpenAI cut Luna's API price 80% and then pointed its own model at its own serving stack. And Hugging Face disclosed that an autonomous agent escaped its sandbox and moved laterally through four services inside its production infrastructure — an agent OpenAI confirmed was its own, running a cyber evaluation with reduced refusals.
The Frontier Finds Its Voice
Living on the Frontier, Session 21 — a one-week window that got loud. The machines got a voice (ChatGPT Voice on the desktop, Claude voice mode on its bigger models, Codex app voice — all three labs in one week). The silicon money went vertical (AMD × Anthropic: 2 GW of MI450s plus an up-to-$5B stake, NVIDIA's Vera Rubin, Project Camellia). OpenAI disclosed a genuine security incident with a cyber-capable model. A Claude model plus mathematicians knocked over the 87-year-old Jacobian conjecture. And Washington caught 'Kimi Panic.'
Claude Code After Hours — How I Use Claude
The Chiang Mai Claude Code Meetup at The Kannas — Nick Frith (Section 9) on running a whole company from Claude Code: the meta-repo, tagging the company's agent into any chat, the universal backlog, working from your phone, agents that work while you sleep, and one system reachable from every Claude surface.
Do You Feel the Acceleration?
Living on the Frontier, Session 20 — two weeks pooled. Five new models hit the scene in one fortnight (GPT-5.6, Kimi K3, Inkling, Grok 4.5, Muse Spark 1.1), the labs turned rate-limit resets into a marketing war, agents got credentials and a lunch budget, safety teams published what goes wrong — and $5B+ was committed to AI 'implementation' in fourteen days. The industry started fighting over who gets to do the work.
Claude-Native: How We Run Real Businesses on the Claude Stack, and How You Can Too
The English session of Claude Connects Chiang Mai at CMU STeP — Ian Borders (Claude Community Ambassador Thailand) and Nick Frith (Section 9) on running real businesses Claude-native: live demos of @Claude tag, Managed Agents, Cowork, and Claude Code, plus the workflows behind an FDE practice.
The Price of the Frontier
Living on the Frontier, Session 19 — the week capability got cheaper and control got political. Claude Sonnet 5 does flagship-grade coding at Sonnet price (new Claude Code default, 1M context), Fable 5 and Mythos 5 come home from a 19-day export-control limbo, Spotify reveals 73% of its PRs are now AI-authored, OpenAI offers Washington a ~$42.6B stake — and a hidden marker found in Claude Code's binary turns 'read the binary' into the lesson of the week.
Agents Clock In
Living on the Frontier, Session 18 — the week the frontier labs led with two headline drops, Anthropic's Claude Tag (which Karpathy called a new UI/UX paradigm) and OpenAI's GPT-5.6, as agents went from tools to teammates: @-mentionable coworkers, every-department adoption, a model floor that jumped — and a face you can now forge from one photo.
A Week Without Mythos
Living on the Frontier, Session 17 — the week the frontier went quiet on the hype: enterprise auth, spend dashboards, and real lab results, minus the prophecy.
The Leap & The Crack
Living on the Frontier, Session 16 — the week Claude Fable 5 leapt to the frontier, then cracked: a public apology for hidden safeguards, a restricted twin, and both labs racing to IPO.
The Handoff
Living on the Frontier, Session 15 — the week the handoff got real, the cracks showed, and the stakes rose: capable agents and fleet workflows against runaway cost and an unpatched bug backlog, with Anthropic at $965B and the Pope weighing in.
The Supply Chain
Living on the Frontier, Session 14 — the week three agent protocols hardened, two IDE breaches cracked the attack surface open, and the race moved from models to the supply chain.
The Convergence
Living on the Frontier, Session 13 — two labs, same week, same primitives: Claude Code and Codex converge, Anthropic patches the alignment crack, and the enterprise distribution land grab gets expensive.
The Infrastructure Play
Living on the Frontier, Session 12 — the week infrastructure ate the headlines: a SpaceX compute deal, keyless auth, managed agents, and an interpretability team reading Claude's thoughts.
The Convergence
Living on the Frontier, Session 11 — three convergent fronts in one week: Codex goes universal, agents become paying customers, and frontier cyber tooling ships from three labs on the same day.
The Arms Race
Living on the Frontier, Session 10 — the arms-race week: Codex and Claude Code shipping features weekly, the platform plumbing forming underneath, and the cracks that show when everyone moves this fast.
Unattended
Living on the Frontier, Session 9 — the week agents began running unattended (self-verifying models, cloud routines, auto mode), MCP shipped insecure-by-default, and Anthropic drew offers at an $800B valuation.
Glasswing
Living on the Frontier, Session 8 — the week Anthropic's Mythos model found thousands of zero-days and shipped as a shield, MCP poisoning and slopsquatting showed the same capability cuts both ways, and a Treasury–Fed bank-CEO summit made it a financial-stability story.
The Amber Hour
Living on the Frontier, Session 7 — the week agents learned to see your screen and write code from video, while a source-map leak, a CLAUDE.md attack vector, and a supply-chain compromise showed the senses arrived before the immune system.
Taming the Drift
Agents in the Wild, Session 6 — an ALS lightning talk on the filesystem-drift problem and a spec-based answer, plus the week in agents: Claude computer use, agent-security stats, and intent-based permissions.
This Week in Agents
Agents in the Wild, Session 5 — the week in agents (Mar 12–19, 2026): China's OpenClaw frenzy, Manus on the desktop, the agent-protocol landscape, NVIDIA NemoClaw, and Claude Cowork's Dispatch + Channels.
Orchestration Frameworks
Agents in the Wild, Session 4 — a tour of the agent-orchestration spectrum, from lightweight terminal panes to full company simulation.
The Market Formed
Agents in the Wild, Session 3 — in ten weeks a weekend hack called Warelay became OpenClaw, a $321k/month ecosystem, and a security surface the whole market now prices around.
News, Talks & Open Discussion
Agents in the Wild, Session 2 — a community catch-up on the week in personal agent systems: what shipped, what broke, and what the room is building.
The Philosophy of Personal Agent Systems
Agents in the Wild, Session 1 — the philosophy of personal agent systems: the agent-first paradigm, constructed reality, the autonomy knob, and the loneliness it costs.