Daily AI Engineering Brief

What Happened

Production agent systems are moving from prompt-based guardrails to infrastructure-level controls. Security teams are discovering that RAG permissions must filter at retrieval time, not in LLM output. Enterprise deployments show agents achieving 96% automation in regulated healthcare by encoding institutional knowledge into orchestration layers. Meanwhile, new tooling addresses core agent challenges: deterministic web scraping converts legacy sites into MCP servers, iterative learning loops build knowledge graphs without retraining, and eval-driven optimization replaces manual prompt tuning. A new memory architecture for coding agents treats context selection as a cost-benefit problem across successive tasks.

Why It Matters

These developments expose a maturation pattern: early agent systems relied on prompts for control, while production systems are building proper infrastructure. RAG permission filtering shows that security cannot be an afterthought—asking an LLM to “un-know” confidential data already in its context window is probabilistic at best. Natera’s 96.2% automation rate in healthcare demonstrates that regulated environments demand deterministic tool boundaries and explicit orchestration, not just better prompts. The shift from browser automation to deterministic HTTP scraping reflects similar thinking: agents need reliable APIs, not brittle pixel-level control.

Key Trends

Infrastructure Over Prompts: Production systems are moving critical controls out of LLM context. RAG security requires filtering documents before embedding, not filtering outputs after generation. Natera’s architecture uses hard-coded rules and orchestration layers alongside RAG, explicitly choosing where to trust LLM judgment versus deterministic logic.

Incremental Learning Without Retraining: Kapso’s knowledge graph approach and CoReMem’s prospective valuation both address the same problem: agents need to compound understanding across tasks without full retraining cycles. Kapso runs experiments and updates graphs incrementally. CoReMem measures which memory segments will pay back their token costs on future tasks.

Measurement-Driven Development: Harness Tuner replaces manual prompt iteration with automated optimization loops using train/val/test splits. CoReMem measures memory value by running tasks twice—with and without each segment—to quantify utility. Both tools treat agent improvement as an empirical problem requiring proper evaluation infrastructure.

API Extraction from Legacy Systems: Pomerado’s MCP servers convert bank sites into structured APIs by recording workflows once, then generating deterministic HTTP scrapers that self-heal when DOM changes. This matters because most enterprise integration targets will never ship official APIs, and browser automation is too slow and often violates terms of service.

Cost-Aware Context Management: CoReMem explicitly models write, read, and storage costs against uncertain future utility. This represents a shift from “include everything” RAG to selective memory based on measured value, critical as context windows grow and token costs compound across long-running tasks.