The Ralph Loop Goes Mainstream in 2026 — and Brings a Token-Burn Problem
Autonomous coding agents in an endless loop: what Ralph Loop and Loop Engineering really are — and why the bill explodes without hard stops.
Redaktion
AI Concepts
Grundbegriffe und Mechanik moderner Sprachmodelle und KI-Workflows.
-
-
Token Tsunamis: why agentic AI systems blow up your costs
2026 industry data: agentic AI burns 5–30x more tokens than a chat. What that means for multi-agent budgets.
Redaktion -
Anthropic: Claude Writes Over 80% of Its Own Code — and Calls for a Pause Button
Anthropic discloses the numbers: Claude writes >80% of its production code. At the same time, the company argues for the option to pause frontier development.
Redaktion -
Agent Orchestration 2026: LangGraph Becomes the Default, Microsoft Consolidates Onto One Framework
LangGraph overtakes CrewAI in stars, Microsoft Agent Framework 1.0 is GA. The framework market is reordering — what that means for your choice.
Redaktion - ki-konzepte Agenten
Anthropic Threat Report: AI-Using Hackers Rise From 33 to 56 Percent
Anthropic's first major threat report: the share of high-risk actors using AI nearly doubled. Plus one largely autonomous agent attack.
Redaktion -
Claude Code Dynamic Workflows: the agent swarm enters research preview
Anthropic opens the research preview of Dynamic Workflows in Claude Code: up to 1,000 subagents per run that adversarially review each other.
Redaktion -
Claude Agent Skills — how a SKILL.md file replaces the custom agent
Anthropic establishes Agent Skills as an open standard. One Markdown file now does the job that used to require a full custom agent.
Redaktion -
Google Gemini Enterprise Agent Platform — Vertex AI gets a real agent stack
GA since April 22, 2026: Google fuses Vertex AI, Agent Builder, and Memory Bank into the Gemini Enterprise Agent Platform — one stack for the lot.
Redaktion - ki-konzepte Evaluation
Apple's 'Illusion of Thinking' — what's left of it one year later?
A year ago Apple published its reasoning critique 'Illusion of Thinking'. We look at which findings held up and which were refuted.
Redaktion - ki-konzepte Training & Fine-Tuning
MIT researchers double the training speed of reasoning models
Adaptive Drafter uses otherwise-idle GPUs during the RL training phase and speeds up reasoning training by 70 to 210 percent.
Redaktion - ki-konzepte Training & Fine-Tuning
On-Policy Distillation: 30 times cheaper to train your own reasoning model
Thinking Machines Lab shows how small models learn reasoning without full RL. Qwen3-8B hits 70 percent on AIME'24 at one-thirtieth the compute cost.
Redaktion -
OpenAI AgentKit — drag-and-drop replaces the agent build-from-scratch
OpenAI bundles Agent Builder, ChatKit, and Evals. Visual workflows replace boilerplate — ChatKit and Evals have been GA since DevDay 2025.
Redaktion -
Code with Claude 2026: multi-agent orchestration, routines and dreaming
At its devcon, Anthropic announced multi-agent fleets, outcomes-based control and a self-improvement loop for managed agents.
Redaktion -
A2A Protocol After One Year: 150+ Organizations, In Every Major Cloud, First Stable Spec
At its one-year mark the A2A protocol counts 150+ organizations, ships in every major cloud, and reaches its first stable 1.0 spec for agent interop.
Redaktion