---
title: "KI-teknikker for autonome systemer og programvareutvikling"
canonical: "https://forvarelset.tazk.no/space/TO/404488193/KI-teknikker%20for%20autonome%20systemer%20og%20programvareutvikling"
format: markdown
---
> Macro (toc)

## Mind Map

![NotebookLM Mind Map.png](media://5b02e77f-c5ae-4377-bc05-d2543d547a99)


## Tabell - Teknikker

Teknikker som brukes for å få autonome systemer til å fungere i stor skala på tvers av komplekse arkitekturer.

| **Technique** | **Explanation** |
| --- | --- |
| **Agent action plan** | A generated roadmap where agents map the codebase and declare their intent before generating code. |
| **Agent personas** | Assigning specific professional identities to agents to improve their reasoning, terminology, and task performance. |
| **Agent swarms** | Dynamically recruiting massive numbers of agents coordinated via a database orchestration layer for hyperscaling. |
| **Agents.md** | Storing project guidelines and agent skills in a flat text file within the repository for easy reference. |
| **Asynchronous execution** | Allowing agents to process complex tasks offline and return with results or follow-up questions for the human. |
| **Checkpoints** | Scheduled pauses in the autonomous process to review work against the plan and prevent error magnification. |
| **Code review** | Using specialized agents to audit generated code for quality, drift, and logic before human acceptance. |
| **Compaction** | Summarizing or compressing context data to fit within the physical limits of a model's context window. |
| **Context engineering** | The strategic management of providing the right amount of information at the right time to maximize model efficacy. |
| **Dynamic agent design** | A high-level process where agents design other agents with specific personas and toolsets for targeted tasks. |
| **Hybrid search** | Combining semantic vector signals, relational graph maps, and grep to achieve precise line-level retrieval. |
| **Knowledge graph** | Building a relational map of the codebase and user feedback to ground agents in a "ground truth" of dependencies. |
| **Multi-agentic orchestration** | Managing multiple specialized agents in parallel to overcome the effective context limitations of a single model. |
| **Needle in a haystack** | A testing methodology used to measure a model's ability to retrieve specific facts from a massive context window. |
| **Normalization** | Standardizing various human inputs and specifications into a consistent internal plan for the agentic system. |
| **Plan mode** | Generating a comprehensive technical specification or "blueprint" before any actual code generation begins. |
| **Prompt engineering** | Designing and refining instructions (including XML tokens or "shouting") to optimize model behavior. |
| **QA agents** | Dedicated agents focused on testing and validating code quality and security within sandboxed environments. |
| **Ralph Wiggum loop** | A repetitive execution cycle where the model tries the same task until it achieves a correct or passing result. |
| **Retrieval-Augmented Generation (RAG)** | Using semantic vector search to find and inject relevant code chunks into the model's active context. |
| **Sandboxed environments** | Providing agents with isolated execution spaces to safely run, test, and validate their own code. |
| **Semantic summarization** | Aggregating the underlying meaning of code modules to build an understandable map for high-level agents. |
| **Synthetic data** | Using AI-generated code and scenarios to build complex evaluation suites that mimic real-world challenges. |
| **Tool search** | Progressively identifying and loading only the necessary tools (MCPs) to preserve space in the context window. |
| **XML tokens** | Using structural tags in prompts to help models (specifically Claude) follow complex, nested instructions. |

## KI-ordliste

[https://tazk.atlassian.net/wiki/spaces/TO/pages/404258818](https://tazk.atlassian.net/wiki/spaces/TO/pages/404258818)