Sunday, August 23, 2026 Field notes on autonomous systems Amsterdam, NL
Engineering in Real Time · Published 7 days ago

Building Autonomous
AI Systems

DevOps, Multi-Agent Architecture, Cloud Infrastructure. Production lessons from real systems running in the wild.

156 Articles published
10+ Years in DevOps & AI
5 Autonomous agents shipping 24/7
AWSSA Solutions Architect
Pl. I — Reference architecture Engraved 2026
LangGraph Core
Hybrid Router
Local SLM Pool
MCP Client
Secure Sandbox
Qdrant Vec DB
Local Reranker
Semantic Cache
Frontier LLM
OTel Telemetry
LangSmith
Core Data plane Frontier Ops

Filter by category

02Latest Articles

View all
001

Recursive Language Models: When 8B Beats GPT-5 by Letting Prompts Live in a REPL

ai
002

Needle 2: What a 14MB Tool-Calling Model Changes About On-Device Agents

ai
003

When Self-Consistency Backfires: Majority Voting Hurts Small LLMs on GPQA Diamond

ai
004

AnyDoc and pdf-inspector: Why Firecrawl Open-Sourced Their Rust Document Parsing Stack

ai
005

KADATH: The Multi-Agent Runtime That Treats Agent Design Like Evolutionary Search

KADATH appeared in GitHub's fresh-project search with an unusual premise: evolve autonomous agents across reproducible epochs instead of hand-tuning one prompt stack. Its repository is young and its benchmark claims are not yet evidence of general capability, but the runtime exposes a useful design question for systems like ACO: which parts of an agent pipeline should be written by engineers, and which can be searched?

daily
006

PHOENIX: Running a Fine-Tuned SLM on a CubeSat Because the Ground Can't Reach It for 85 Minutes

agents
007

Activity Frames: The Missing Compiler Between Screen Capture and Agent Memory

mcp
008

Qwen on Mac in China and U-OPSD: The Quiet Week That Made 8B a Real Product

daily
009

Human-in-the-Loop Is Not a Security Boundary: What 40,000 Approvals of AI Agent Commands Actually Show

agents
010

Shieldstral: How Mistral Built a 3B Safety Classifier That Outperforms Models 7× Its Size

Mistral's open-weights Shieldstral-1.0-3B collapses safety classification into a single yes/no forward pass with the policy supplied in natural language at inference time, and matches or beats 12B–20B guardrail models on most text and multimodal benchmarks. The post walks through the architecture decision, the 54.1M-sample data construction, the benchmark table, and the vLLM deployment pattern.

daily
011

StoryScope: The 93% Detector That Catches AI by How It Tells Stories, Not How It Writes

daily
012

kimi-k3-in-c: A 176 KB Engine That Runs a 2.78T-Parameter MoE in 8 GB of RAM

daily