Descriptions:
Nate B Jones delivers a technical deep dive into what he calls ‘agentic context engineering’ — the discipline of managing what information agents see, when they see it, and how it evolves across long-running tasks. The video synthesizes three research papers published in late 2025: Google’s Agent Development Kit (ADK), Anthropic’s Agentic Context Engineering (ACE) paper, and Manus’s paper on long-running agent design, presenting them together as the first coherent blueprint for production-grade agent memory systems.
The core argument is that longer context windows have not solved the agent memory problem — they’ve intensified it. Naively appending conversation history causes attention dilution, log bloat, and performance degradation on extended tasks. The nine principles Jones extracts from the papers include treating context as a compiled view rather than a transcript (dynamically assembling only what’s relevant per LLM call), building tiered memory models that separate working context from session logs from durable memory, schema-driven summarization of tool results rather than raw output injection, and using sub-agents to isolate scope rather than mimic organizational hierarchies.
Practical engineering payoffs are quantified: structuring prompts around stable prefixes to maximize cache reuse can reduce per-step latency from ~200ms to ~20ms, dramatically cutting both cost and response time in multi-step agent loops. The video also covers Anthropic’s approach to evolving agent strategies through execution feedback — allowing agents to refine their own instructions incrementally rather than requiring human prompt re-engineering. Aimed at engineers and technical architects building production agent systems with frameworks like Google ADK or the Anthropic API.
📺 Source: AI News & Strategy Daily | Nate B Jones · Published December 09, 2025
🏷️ Format: Deep Dive
