---
title: "KI-ordliste"
canonical: "https://forvarelset.tazk.no/space/TO/404258818/KI-ordliste"
format: markdown
---
| **Concept** | **Definition / Explanation** |
| --- | --- |
| **Agent** | Specialized AI units designed to perform specific tasks, such as mapping codebases, writing code, or performing quality assurance. |
| **Agent action plan** | A strategic plan generated by agents that maps the codebase and declares the intent for autonomous code generation. |
| **Agent personas** | The practice of assigning a specific professional identity to an agent to influence its reasoning process and terminology. |
| **agents.md** | A flat file (simple file) used to store specific guidelines and rules for AI agents within a project's codebase. |
| **AI assisted development** | The use of AI tools like co-pilots and IDE assistants to support human developers, rather than replacing the workflow entirely. |
| **Autonomous development** | The end-to-end automation of software engineering where AI generates, tests, and validates code without constant human intervention. |
| **BERT** | A transformer-based model used for natural language processing tasks, such as language translation. |
| **Blitzy** | An enterprise-scale platform for autonomous software development that utilizes agent swarms and knowledge graphs. |
| **Claude Code** | An AI tool from Anthropic designed to assist with coding and development tasks. |
| **CLI** | Command Line Interface; a text-based interface where AI assistants can be embedded for developer use. |
| **Codex** | An AI model or tool used for code generation and reviewing existing codebases. |
| **Compaction** | A process where a model summarizes or reduces information to fit within the constraints of its context window. |
| **Context engineering** | The practice of optimizing the information provided to an agent to maximize efficiency and avoid overloading the model. |
| **Context window** | The maximum amount of data (tokens) that a model can process and "keep in mind" during a single request. |
| **Cyclomatic complexity** | A technical metric used to measure the complexity of code, which directly impacts its maintainability. |
| **Effective context window** | The actual limit of information a model can handle while still maintaining peak performance and accuracy. |
| **Eval (Evaluations)** | Internal or public frameworks used to systematically measure and compare the performance of AI models and agents. |
| **GANs** | Generative Adversarial Networks; a class of machine learning frameworks used for generative tasks. |
| **Gemini** | A family of large language models developed by Google. |
| **GPT** | Generative Pre-trained Transformer; a family of language models developed by OpenAI. |
| **Graph database** | A database that maps relational structures and dependencies in a codebase to ground agents in a "ground truth." |
| **Hybrid search** | A technique combining semantic vector search for direction and graph/grep search for exact line-level precision. |
| **IDE** | Integrated Development Environment; the primary software application used by developers where AI assistants are often integrated. |
| **Knowledge Graph** | A self-reinforcing structure that stores relationships, metadata, and user feedback to improve autonomous development over time. |
| **LLM** | Large Language Model; the underlying artificial intelligence that powers agents and generative systems. |
| **MCP (Model Context Protocol)** | A protocol that allows AI models to connect to external tools and integrations, such as browsers or databases. |
| **Multi-agentic** | Systems that utilize multiple specialized agents working together to solve complex, high-level tasks. |
| **Needle in a haystack** | A benchmark test measuring a model's ability to retrieve a specific piece of information from a massive context window. |
| **Opus** | A high-performance model from Anthropic, part of the Claude 3 family of models. |
| **Orchestration layer** | The system component that recruits and coordinates multiple agents, often utilizing a database for scaling. |
| **Prompt engineering** | The process of designing and refining instructions (prompts) to guide the behavior and output of AI models. |
| **PR (Pull Request)** | The final output of a development task, containing the code changes ready for review and acceptance. |
| **RAG (Retrieval-Augmented Generation)** | A technique that retrieves relevant information from external sources to provide the model with better context. |
| **Swarms** | Dynamically recruited groups of thousands of agents that work in parallel to achieve hyperscaling in development. |
| **SWE-bench** | A popular leaderboard and evaluation framework used to test AI models on real-world software engineering issues. |
| **Tokens** | The basic units of text (words or characters) that language models use to process and generate information. |
| **Tool calling** | The ability of an AI model to interact with and execute actions using external software tools. |
| **Transformer** | The foundational neural network architecture that enabled the current wave of generative AI and language models. |
| **Vector database** | A database that stores information as mathematical vectors to enable semantic and similarity-based searching. |
| **XML tokens** | Specific tags used to structure prompts, frequently used to improve the performance of models like Claude. |