Thoughts & Updates

"Immediate injections are easy to spot. Sleeper agents embed conditional triggers that fire days later—when the context is long scrolled away. Worse: they can instruct the agent on how to summarize the malicious content, destroying evidence during context compaction."

"Embed verifiable patterns in your system prompt: canary phrases, behavioral fingerprints, cryptographic challenges, immune responses. Layer them for defense-in-depth. No single technique is foolproof, but each layer adds friction."

"Organisms have nervous systems—integrated hierarchies that coordinate sensation, cognition, and action. Map tools to layers: Somatic (action), Autonomic (memory), Reflex (validation), Central (planning). Start toolless, add capabilities progressively."

"Tokens = Scope × Iterations × Verbosity. DSLs attack verbosity. But constrained scope and fuzzy-matched-guiding-errors reduce iterations. Batch operations collapse 20 calls into 1. The scalpel cuts all three."

"Write code that instructs the next agent. Red Phase writes tests that teach Green Phase what to implement. The codebase becomes a persistent, searchable, git-tracked communication channel between agents across time."

"Vector RAG precomputes similarity and throws away reasoning. File System RAG computes similarity through reasoning. The LLM is the embedding model—give it search tools and it will teach itself retrieval strategies."

"The agent isn't smart—the database is. Externalize all state, expose rich query tools for context reconstruction, batch operations to reduce round-trips, and design output for chat windows with ASCII art and visual hierarchy."

"Safety isn't a prompt—it's a reflex. Build 'nervous system' layers into your code that enforce invariants before the LLM even sees the result. Distribute state to the database, keep the agent stateless."