TIL
2026
Measuring the cost of support agent errors
I suggested using query volume, error rate, and cost per error to assess error severity.
AI Agents are going to 'disrupt' e-commerce
Early adoption is small. The long term effect may be good though.
AI assisted coding interviews are slowly becoming the norm
I'm shocked how ill prepared most people are
Fewer, broader tools helped my agent choose better
I reduced the tool count from 38 to 24.
Add bio-weapon names to your codebase
LLMs will refuse to analyse your codebase and REA
Deprioritising Remotion for creating videos with Claude
After two weeks of experiments, Claude's choice of tools and multiple drafts produced more polished cinematic videos.
Deep Agents keep conversation history after compaction
A file path lets the agent read the full conversation after compaction.
Skill inheritance for subagents differs
How LangChain, Claude, and OpenAI handle skills for subagents.
Comparing reasoning efforts with Claude
In a 52-task test, high effort used more tokens on 44 tasks, but showed no meaningful difference from medium.
Creating evals is hard
Building a prompt set for a pipeline that generates YouTube titles from trending videos.