New in Claude Code: your sessions can now message each other. Instead of having to re-explain yourself in another session, you can now tell Claude to do it. It sends a summary (not your history or files), and the other session picks it up mid-task.

automode is much safer than any other permission system out there, especially reviewing them yourself excited to announce we're rolling it out to everyone by default, with no overhead cost for the classifier

Opus 5 is a shocking step up from Fable 5 and 5.6 Sol for spreadsheets Almost couldn’t believe our evals at first A step function increase in intelligence, while as efficient as Sol Feels like another Opus 4.5 moment for knowledge work

what's the difference between a loop and a graph? (marketing edition) both are ways to run an agent, the difference is who decides the path, the agent or you. a loop still starts with you. you set the goal, the brief, and the bar it has to clear. what the agent owns is the path. take writing an SEO article: hand it the brief and it drafts, reads the draft back against that brief, rewrites the weak parts, checks again, and keeps circling until it clears the bar. the one thing you did not write is the step-by-step it took to get there. a graph is you drawing the steps and the routes between them ahead of time. same article, but now you set the map: research the keyword and the competitors ranking for it. draft from what you find. score that draft against your rubric. if it clears, add the internal links and publish. if it misses, back to the draft. the agent still decides how to handle each step, it just travels the routes you laid down. the shape of this has a name, a state machine. every node is a state the work can be in, and a check at each one decides where it goes next, forward when it clears or back to an earlier node when it misses. if you have built a workflow in n8n, you have already drawn one. nodes wired together, branches that fire on a condition, a step that loops until it clears, that picture is a graph. an agent graph is the same shape, the nodes hold agents doing the work instead of single api calls. the way I think about it, a graph is a map of loops and checkpoints. some nodes run once, others are their own loop where the agent works something out, and the checkpoints between them read the result and route the work. you keep laying down nodes and checkpoints until the map reliably gives you the output you want. the vault accelerator I run at my agency is one of these maps, 3 sessions that hand off in a fixed order: > research session: reads our company brain and past campaign results, pulls in competitor and market context, and builds the cohort we go after > landing page session: takes that research and builds the page from it > content session: uses the research and the page to write the copy, illustrations, and slides for the live sessions we run inside the content session runs a loop, a critic scores each draft against a rubric and sends it back until it clears the bar. that is one node on the map, the checkpoints between the sessions carry the work from one to the next a graph earns its extra setup on anything you run every week: > validation gates the work cannot skip > a fixed set of routes the job can take > a clear failure point, you see the exact step something broke on a loop on its own is enough for the work you only do once, where you don't know the path yet, let the agent find it. graphs earn their place on the jobs you repeat, the content pipeline, the SEO and AEO funnel step by step, the vault accelerator once the map works you reuse it, feed it the next cohort and the whole pipeline runs again past the loop, the next thing you design is the map it runs inside.

Your ChatGPT Work agent can now use websites that require you to sign in. Take over the cloud browser to log in, then let your agent continue the task. Your login persists across sessions, so you only have to sign in once.

Many people think any given ML project is 99% training. In reality, it’s 50% evaluation, 40% data cleaning, 8% integration, and 2% training. The first two set the noise floor for learning. No ML magic matters; the model cannot lower the noise floor, as that’s the optimal bound of Shannon encoding of your data. Thus, not a single day goes by without me thinking about ontology. Even the old labels have to be constantly reviewed.