Your AI Agent Scores 80%. Which Tasks Can It Repeat?
A small Python evaluator shows why task-level outcomes and clean trial resets belong in the same agent test harness.
Sep 16, 20265 min read
Command Palette
Search for a command to run...
Latest articles
MetaRoCE: Why a Million-GPU Cluster Must Tolerate Loss
Meta's stated shift toward loss and reordering tolerance points to the missing contract in many distributed training stacks.
Aug 26, 20267 min readAgentic Git Hosting Needs a Review Queue, Not Faster Pull Requests
Cursor Origin points at a real bottleneck: once agents create changes cheaply, a forge has to ration reviewer attention with evidence and explicit merge lanes.
Aug 19, 20267 min readDiffusion LLM Caching Is a Scheduler Problem, Not a Cache Problem
LaCache makes the case for exact reuse, but the serving win comes from coordinating cache identity and precision at the same boundary.
Jul 22, 20267 min read1A 3.9 GB Model Does Not Make a Phone Agent: Runtime Memory Does
A 5-7 MB WASM embedding model changes the tradeoff: for many docs and support flows, the right search layer now lives in the browser, not behind another inference API.
Jul 15, 20267 min readOpenRouter Fusion: the synthesis step does the work, not the panel
Fusing cheaper models matched Claude Fable 5 at half the cost. OpenRouter's own numbers say about 75% of the gain came from one rewrite pass. Here is how to capture it without paying for six frontier calls.
Jun 17, 20267 min read2The real cost of running agents is the KV cache, not the tokens
Open-weight models keep getting cheaper to serve. The reason is memory, and a wave of KV-sharing and compressed-attention tricks is quietly rewriting agent economics.
Jun 15, 20266 min read4