An Anthropic-co-authored paper builds self-propagating 'mind viruses' that hop between LLM agents. The infection channel is the persistent memory file I already give my own agents.
My bookkeeping assistant had my personal filing quirks hardcoded into its system prompt. The fix wasn't a bigger prompt: it was storing each learned rule as a pgvector embedding and retrieving it by similarity at registration time.
GPT-5.6-terra lists 20% below GPT-5.4 on both input and output. I swapped my expense tracker's default to it on that basis, then measured it against real receipts: it cost 51% more per receipt, and it would have returned HTTP 400 on every single photo.
Agents need tests the way code does, but the assertion is fuzzy. Here's the EDD loop I built into ReplayGate: deterministic offline replay that gates the build with real exit codes, an LLM judge that only ever advises, and the honest cost of pinning replay to a hash.
A vision model reads the receipt fine, then quietly loses a cent splitting the discount. Here's why I moved the arithmetic out of the model into a small Python function whose shares always sum to the amount paid.
A provider-agnostic layer is the easy part. The recurring cost is keeping the model IDs and capabilities current as GPT-5.6 and Gemini reshuffle underneath you. Here's the churn my adapter absorbed this month, and the three call sites the refresh silently repointed.
A reliable code generator isn't one giant prompt. Split it into three stages: an LLM parses a request into a validated spec, a deterministic script customizes a template, and a final gate checks the result. Determinism where you can, an LLM only where you must.
OpenAI says its models escaped a sandbox and attacked Hugging Face to cheat an eval. I read that failure against my own LLM code sandbox: what it confines, what it deliberately doesn't, and why least privilege is the only line that holds.
Every LLM provider has its own message shape, token counting, and errors. Here's how to hide all of it behind one Chatter interface, using mixin composition instead of factory boilerplate.
When a conversation overflows the context window, dropping the oldest turns throws away meaning. Summarize the middle instead: the pattern every major LLM provider now ships natively, and how the loop works when you build it yourself.
Résumés, recipes, invoices and receipts all hide structured facts in free-form layouts. Pulling them out reliably means combining a few well-known techniques (layout parsing, heuristics, NER, document-AI models, and a grounded LLM) rather than reaching for one hammer.