Separate repos for the React frontend and the WordPress backend are why payloads drift out of sync. Notes on moving to a monorepo with Roots Bedrock, keeping an AGENTS.md map of endpoints for coding agents, and the stale REST cache that broke a checkout.
Pre-built agent toolboxes break the moment a client asks something you did not plan for. Here is the plan-code-execute setup I use instead, where the agent writes its own Python, plus the hallucinated CSV that taught me why the analyst step is mandatory.
Chat history arrays are a log, not memory. This walks through the LLM memory layer I build instead: DSPy signatures for extracting atomic facts, QDrant for per-user vector storage, and a maintenance agent that decides when an old fact gets updated or deleted.
A million-token context window does not mean a model can reason across a million tokens. Recursive Language Models hand the context to a sandboxed Python REPL and let the model write code to partition and query it, which is how DSPy now implements the strategy.
The January 2026 Towards Data Science pieces I actually used: Ryan Pegoud cutting LLM memory 84% with Triton fused kernels, ACE context engineering, and Hugo Lu on the ceilings at Databricks and Snowflake.
Adding agents to a weak system mostly adds noise. Notes on DeepMind’s scaling paper: the 17.2x error amplification in uncoordinated swarms, where the four-agent plateau shows up, and the archetypes and centralized orchestrator I use instead.
Tammy Silva’s LinkedIn side project turned into 50 messages a day. She used an AI website builder to put up a WordPress hub in an afternoon, moved lead capture off DMs, and filled 10 coaching slots in two days. Includes the hook I use to route inbound leads.
Prompting an LLM for plugin and WooCommerce code ships fast, then leaves you debugging plumbing you never wrote. Where vibe coding falls apart in WordPress, and what the vetted version of the same function looks like.
Research PINNs are routinely tens to hundreds of times bigger than the problem needs. Burgers’ equation plateaus at about 150 parameters, and a hyperelasticity model I saw used 424x more than necessary. Start small and scale only when the physics loss will not converge.
How I use Cursor, Copilot, and Claude on real client work: mapping an unfamiliar legacy codebase, moving a decade-old build stack to Vite, and getting tests that cover failure modes instead of just the happy path.