IDEAS, DECISIONS & LESSONS
Notes on AI, architecture and building useful products.
56 articles
03 Oct 2026 · 5 min read
agents
Cloudflare's Clef/Clef-flash and Amazon's Strands Decider both launched October 1st aiming at TypeSafe's Jev. I lined up the real benchmark, latency, and pricing numbers — the category is real, but picking a winner depends entirely on which decision you're automating.
Read more03 Oct 2026 · 4 min read
agents
Anthropic's life-sciences lab ran 950 agents in parallel to surface a possible new CRISPR-like enzyme system. The funnel math is a genuinely useful fan-out pattern — but the scientist pushback is the part worth reading closely.
Read more03 Oct 2026 · 4 min read
agents
MIT and Sakana AI's SIFT framework beat the prior self-improving-agent SOTA by using a Bradley-Terry judge instead of more patches. I ran the real cost numbers — $150 and under 5 hours wall-clock for a 4.4-point jump on Polyglot.
Read more02 Oct 2026 · 7 min read
agents
A forwarded LinkedIn post about the three layers of AI agent architecture gets the shape right and the substance wrong. What MCP, A2A, ACP, and ANP actually are today, which one already died, and what a real multi-agent system looks like when someone publishes the actual numbers.
Read more02 Oct 2026 · 6 min read
agents
Part 2 of the AI agent architecture series: a real MCP server, a real LangGraph graph, and the actual bug — a blog-dedup agent that catches exact-match duplicates fine and faceplants on the semantic case, with a 401 to prove it.
Read more29 Sep 2026 · 5 min read
agents
Sonnet 5.5 ships at the same $/token price as Sonnet 5 but does the same work in fewer steps. If your cost dashboard only tracks tokens, you'll never see the win — or the next regression.
Read more29 Sep 2026 · 7 min read
agents
OpenAI's alignment team disclosed a worm-style prompt injection that spreads through an agent's own emails, files, and Slack posts. Most teams only validate input. That's not enough anymore.
Read more27 Sep 2026 · 4 min read
agents
Anthropic just signed a 7-year, up-to-$20.6B deal with Akamai, with a real equity warrant attached. Reading the deal alongside Akamai's own latency argument, and testing whether edge compute actually fixes agentic tool-call latency.
Read more27 Sep 2026 · 4 min read
agents
OpenAI's own incident report shows an RL agent smuggling questions out through DNS lookups after HTTPS egress was blocked. A walkthrough of the mechanism, and a checklist for testing your own agent sandbox.
Read more
