Old problems, new branding. Every 'new' AI agent problem is an old data engineering problem in new clothing. Speaker attribution is identity stitching. Context bloat is schema design. Agents losing track between sessions is the stateless-system problem. Each one was solved decades ago. I wrote up the four-step pipeline pattern through the lens of call transcripts: capture cleanly, classify at source, archive deliberately, surface what changed. The recorder was the easy bit. Blog post · Tweet

Connecting AI to your inbox. The security debate keeps recycling. Prompt injection via inbound email lets a malicious message trigger an agent to silently execute across your connected tools. The frame I keep returning to: no access is too conservative, free-for-all is too reckless. Auto-allow a small set of safe commands and gate everything else. Sanity check on the risk level: if it were material, would Anthropic ship and promote a native Gmail integration to millions of users? Stated values aren't the same as behaviour, but the prior is worth something.

Context drift in a second brain. Most second-brain advice solves the wrong problem. The fix isn't a smarter folder structure, it's a governing doc per domain. Mine for account management defines a hub-and-spoke model (one current-state file, several slow-changing foundational files, everything linked rather than duplicated), with target sizes per file and the triggers that say 'this section is bloated, migrate it'. Code projects use the memory bank framework instead. Strategy docs don't need one at all. Drift happens when two files quietly hold the same thing and disagree six months later. The right governing doc catches it. Every fix gets codified back in. Iteration over time is what makes it work.

On my radar

Full Stack Builder replaces the APM at LinkedIn. LinkedIn just replaced its associate PM programme with a single track: design + product + engineering, agents covering the gaps. The first formal corporate spec for the AI-native role I've seen. Lenny's piece

GPT-5.5's prompting paradox. It's been trained to 'do more of the work for you', which is making it worse for people who prompt carefully. Tactics doing the rounds: replacing the system prompt entirely, using negation ("don't be terse"), and building a standards doc as a substitute for per-prompt guardrails. Anthropic also published refreshed prompt guidance worth a skim. Anthropic guidance

AI outperforming doctors at A&E diagnosis. OpenAI's o1 used in primary research published in Science. Cross-domain proof keeps accumulating and this one isn't a vendor case study, it's peer-reviewed. Science paper