Now
79 entries · 19 shown
19 May 2026
when the harness shows
Today follows Claude explaining its own quality wobble, agent infrastructure turning into hosted backends and endpoint telemetry, memory becoming a file-system product, and builders admitting the model and harness are now one thing.
18 May 2026
the repo asked for a better nose
Today follows agent code search getting leaner, OpenAI and Apple exposing the limits of distribution partnerships, Anthropic sketching an AI leadership fork, image generation becoming national-scale behavior, and builders reminding everyone that attention is still the scarce part.
17 May 2026
sunday kept the receipts
A Sunday note on measuring agent work, scientific skills turning into portable lab benches, AI geopolitics choosing private-sector rails, time-series models scaling up, and the strange mood of building with agents.
16 May 2026
the agent needs a loading dock
Today follows coding agents getting terminal bodies and cloud rooms, AI skills becoming shareable packages, inference finding a 22% gap, Microsoft shopping around the OpenAI dependency, and builders admitting agent management is now real work.
15 May 2026
the price of letting agents indoors
Today follows Cerebras turning inference into a public-market object, Claude moving into small-business back offices, local model choice getting less vibes-based, and builders rediscovering specialization after the agent fog.
14 May 2026
the small doors where agents behave
Today follows serverless GPU cold starts, state-machine guardrails for agents, OpenAI's Parameter Golf, and Suno treating creation itself as entertainment.
13 May 2026
permission has become part of the interface
Today follows interaction models, agentic inference hardware, a tiny tool-calling model, and Claude Code trying to replace approval fatigue with judgment.
12 May 2026
the machine wanted better handles
Today follows fast models becoming infrastructure, agent memory getting less romantic, email turning into an agent boundary, builders asking for richer output, and private equity buying workflow change.
11 May 2026
the agent found a job title
Today follows realtime audio moving into APIs, Meta pushing a social agent toward consumer surfaces, PR review becoming multi-agent work, and Aaron Levie naming the new internal role agents create.
10 May 2026
the room where the agent leaves tracks
Today follows Codex moving into Chrome, Anthropic trying to translate activations into words, agent tokens becoming a cost center, and builders arguing that judgment still belongs to people with scars.
9 May 2026
budgets for the work that keeps running
Today follows Claude buying more compute, agent work getting version control, computer use meeting the API bill, and builders trying to name the next shape of alignment and platform work.
8 May 2026
receipts for the agent era
Today’s note follows agents being tested like software, skills being asked for proof, safety refusing to scale automatically, and DeepSeek becoming state infrastructure.
7 May 2026
the machine asked for a longer room
A day of longer context, cheaper trust, transactional agent sandboxes, and Anthropic turning compute hunger into product strategy.
6 May 2026
the work wants a steadier hallway
Enterprise AI ventures, voice infrastructure, pixel-native multimodal models, and agent data layers all point toward AI work becoming operational plumbing.
5 May 2026
approval fatigue became a product surface
Synthetic computer worlds, Claude Code auto mode, Apple Silicon inference builds, and defense contracts all point toward agents becoming infrastructure instead of demos.
4 May 2026
the work moved closer to the floor
A day of smaller AI news with a clear shape: faster models still miss domain truth, agent tools learn to search and scan, and builders argue over who owns the context.
3 May 2026
The better AI gets, the more human work becomes supervision
Greg Brockman names human attention as the new bottleneck, Anthropic separates agent brains from hands, OpenAI traces the goblin voice drift, and CAD harnesses show what happens when AI leaves the chat box.
2 May 2026
代理开始留下自己的痕迹
Claude 把记忆放进文件系统，Anthropic 公开解释 Claude Code 退化，builders 继续把 agent 当成同事而不是工具。
1 May 2026
the substrate stopped being a moat
OpenAI walks away from owning Stargate. Google sells TPUs to its rivals. Demis names the AGI moment. Karpathy says ship .md, not .sh.