ConTejas Code What is AI engineering and what do AI engineers even do?
Feb 23, 2026
They break down what AI engineering looks like in 2026 and why reliability, observability, safety, and cost matter. Topics include battling hallucinations, building real-time knowledge with tool-augmented truth, and engineering context packets. The conversation explores agent runtimes, the Model Context Protocol, SKILL.md, and why multi-agent concurrency and device-local runtimes like OpenClaw are rising.
AI Snips
Chapters
Transcript
Episode notes
Assemble A Hierarchical Context Packet For Prompts
- Engineer a context pipeline: rewrite the user's question into retrieval queries, fetch candidate chunks, re-rank, deduplicate, and assemble a hierarchical context packet.
- Include system policy first, then user intent, then facts and recent tool results to keep prompts tight and debuggable.
Treat Retrieved Text As Data Not Instructions
- Prevent prompt injection by labeling retrieved content as data not instructions, sanitizing inputs, and enforcing policy isolation at both application and system-prompt layers.
- Add flags like is_instructions:false in retrieved JSON and explicitly tell the model retrieved text is not executable instructions.
Agents Are Runtimes Combining Model Tools And State
- An agent is a runtime combining a language model, tools, and state; it orchestrates tool calls, context management, and multi-model or summarization strategies.
- Tejas frames agents as environments like a browser for JS, where the runtime executes model-driven workflows and handles tool execution.