Blog

Patterns, comparisons and decision rules for building with LLMs and agents.

15 postsLatest 31 Oct 2026

2026

The Temptation to Build EverythingBuilding a tool is now quicker and more fun than adopting one, which is fine until other people depend on it.
31 OctAI Agents
Your Agent Memory Needs Two ClocksA single timestamp cannot distinguish a late correction from a real-time state change.
26 JulAI Agents, LLM Infrastructure
Delete System-Prompt Rules One Regression at a TimeTie every prompt workaround to a failure case, then test its removal when the model changes.
26 JulAI Agents, LLM Infrastructure
Treat Reasoning Controls Like API MigrationsShip the model, reasoning mode, and history adapter as one versioned protocol.
25 JulAI Agents, LLM Infrastructure
The Tool List Is Not an Authorization BoundaryRemoving a tool from the model's list doesn't revoke the credentials behind it, so check permissions when the call executes.
25 JulAI Agents, LLM Infrastructure
What Pass Rates Hide About Long-Horizon AgentsReport the share of runs that reach each milestone, so you can see the step where an agent starts failing.
25 JulAI Agents
Sliding-Window Prefix Caches Need Physical ProofIn hybrid sliding-window models, the prefix tree can remember tokens whose local KV slots are already gone.
11 JulLLM Infrastructure
Speculative Agents Need a Commit BoundarySpeculative execution cuts agent latency only when prediction and commitment are separate operations.
11 JulAI Agents, LLM Infrastructure
Loop / Harness EngineeringAgent loops turn prompts into one part of a production harness: tools, state, checks, approvals, and exit conditions.
28 JunAI Agents, LLM Infrastructure

2025

2024