Chapter 9.1 — Core concepts (mini-glossary)¶
Just as a quick reference; the full glossary is in glossary.md:
- A2A — an emerging protocol for interoperability between agents; based on Agent Cards and a task lifecycle.
- Agent harness — the engineering around the model that makes the agent operable, safe, testable and observable.
- Agent runtime — the layer that executes the agent's loop.
- Agency — the operational capacity to pursue goals via intermediate decisions.
- AgentOps — practices for operating agents with tools, state and runtime.
- AI gateway — a governance layer between the application and model providers: routing, budget, caching, observability, DLP, policy.
- AWQ — Activation-aware Weight Quantization (quantization that preserves salient channels).
- Limited autonomy — restricted by scopes, permissions, budgets and approvals.
- Context engineering — the discipline of selecting, ordering and structuring context.
- Continuous batching — a scheduling technique that adds new requests to the batch at each step.
- CycloneDX — OWASP's SBOM format; supports ML-BOM.
- Eval harness — continuous evaluation infrastructure.
- GGUF — llama.cpp's file format for quantized models.
- GPTQ — layer-by-layer post-training quantization with a second-order approximation.
- GraphRAG — RAG that uses a graph structure for relations.
- Groundedness — how much the answer is supported by evidence.
- KV cache — cache of attention keys/values; central in decode.
- Matryoshka embeddings — embeddings trained so that prefixes are semantically valid (Kusupati et al., 2022).
- MCP — Model Context Protocol.
- Memory poisoning — contamination of memory with incorrect/malicious data.
- Model routing — dynamic model selection.
- NF4 — 4-bit "NormalFloat", a format used in QLoRA.
- PagedAttention — an algorithm for managing the KV cache in paged blocks (vLLM).
- Policy-as-code — codified, testable and versioned policies.
- Prompt engineering — formulating instructions.
- Prompt injection — input that tries to change the model's behavior.
- QLoRA — fine-tuning LoRA adapters over a 4-bit base (NF4).
- RAG — Retrieval-Augmented Generation.
- SBOM — Software Bill of Materials.
- SLSA — Supply-chain Levels for Software Artifacts.
- Speculative decoding — a fast draft model proposes tokens; the large model verifies in parallel.
- SPDX — the Linux Foundation's SBOM format; ISO/IEC 5962:2021.
- Tool registry — a versioned catalog of tools.
- TTFT — Time To First Token (prefill latency).
- TPOT — Time Per Output Token (latency per generated token).