Seb·Oct 21Search Inside the Model: Built-In Web Search in OpenRouter, OpenAI, Anthropic, Gemini and xAIBuying search from the model provider is the shortest path to citations, and the hardest place to control what gets bought.
Seb·Oct 21Search Tools for Agents in Python and Node.jsThe API is the easy decision. The hard one is what your loop does with the text it gets back.
Seb·Oct 21Prompt Injection Defense in Python and Node.js: The Filter Is the Easy PartA classifier in front of the model is the cheap layer. The one that holds is the layer that decides what the model can do after it has been…
Seb·Oct 13Agent Memory Should Be a Filesystem, Not a Vector IndexOpenViking turns what an agent knows into a directory tree you can open and edit. That is a bigger change than it sounds.A response icon1A response icon1
Seb·Oct 10An Agent That Never Stops: How to Own the ClockNothing wakes itself. The trick is making the thing that wakes it part of the agent, cheap enough to run while nothing is happening.A response icon1A response icon1
Seb·Oct 3Cutting the Bill Is the First Engineering DecisionCheap tokens are the easy half. The expensive half is how the pieces are joined, how often they run, and what stays alive when nobody is…
Seb·Oct 1What 147 Guardrail Decisions Look Like When You Read the LogA guardrail built on a decision model is only as good as the record it leaves. Here is ours, with the failures left in.
Seb·Sep 26Integrating Jev Into Hermes: One Hook, Two Tools and the Numbers From Our Own LogJev never writes code and never approves anything. It answers typed questions, and the framework decides what to do with the answer.
Seb·Sep 25I Replaced a 1.2-Second "Decide What to Click" Step With a 300 ms OneJev is a decision model, and that turned out to be what a browser agent was missing.
InAWS in Plain EnglishbySeb·Sep 24How to Run Your NanoClaw Assistant on DeepSeekThe working config and the cost math.