Context Engineering: Practical Techniques for Improving Agent Quality Today
summary
Claims that an LLM is best treated as LLM = function and tool calls are simply data extraction into unions, with agents implemented as agent = while loop + state. Practical techniques include positioning reasoning tokens early or late, avoiding strict JSON outputs in favor of free text or markdown blocks for better code and token efficiency, and avoiding static few‑shot examples after a healthcare case where an example leaked a rare diagnosis. Emphasis on explicit UX and exception flows for when agents fail.
AI-written summary · about
chapters (12)
- 0:00Context engineering intro
- 1:22Agents as stateful loops
- 3:10Tool calling as data format
- 5:14Parallel tool calls
- 8:43Optimizing with unit tests
- 10:28Reasoning for complex tasks
- 12:57JSON constraints harm outputs
- 15:45Code in markdown blocks
- 17:32Few-shot prompting pitfalls
- 21:49Dynamic few-shot replacement
- 22:55UX design first in agents
- 24:41Error handling with UIs
related talks
01h 00mFrom LLM Theory to Practical Agentic Implementations
41mBuilding LLM Agents – Challenges and Learnings
36mAdvanced Evaluation Techniques for AI Agents
36mLLMs and AI agents demystified
38mNobody Pages the LLM: Engineering Rigour for Vibe Coding
50mIntelligent Agents, the New Frontier of LLMs
56mHerding LLMs: Predictable Outcomes from Unpredictable Agents
01h 43mBuilding LLM Agents (Hackathon)