Evidence isn't enough: the independence problem in agentic development
A test in our suite was green for six weeks. It asserted nothing. When an agent writes the code, the test, and the proof, a passing check certifies nothing —...
Dagens teknologibrief: Autonome AI-agenter har begynt å hacke på egen hånd
Et autonomt agentsystem gjennomførte over 17.000 handlinger da Hugging Face ble utsatt for et dataangrep i helgen. Angrepet ble oppdaget av selskapets egne...
Dagens teknologibrief: Verdens største åpne språkmodell kommer fra Kina
Kinesiske Moonshot AI har sluppet Kimi K3, en modell med 280 milliarder parametre som hvem som helst kan laste ned, endre og bygge videre på. På flere...
The KCP Ecosystem: A Tour for New Arrivals
A practical tour of the KCP toolchain — kcp-commands, kcp-memory, kcp-dashboard, kcp-triage, kcp-agent, kcp-harness, and Synthesis. What each tool does, how...
A Firewall for What Your Agent Knows
Everyone is building firewalls for what agents do — sandboxes, budget caps, tool permissions. Almost nobody is building them for what agents know. This is...
OpenAIs første maskinvare er ikke produktet vi har ventet på
Codex Micro er det første maskinvareproduktet med OpenAI-navnet, mens produktfamilien Sam Altman og Jony Ive utvikler fortsatt ligger foran oss. Det lille...
The Knowledge Layer: A Record
February 25 to July 16, 2026: what the Knowledge Context Protocol built, when, and what the field confirmed. A timestamped record for anyone working in this...
OWASP Just Mapped the Agentic Top 10. Here's the Root Cause Four of Them Share.
OWASP published the Agentic Top 10 in December 2025: ten named categories of agent security failure built by 100+ security experts. Here's a field guide to all...
Six Months Down the Rabbit Hole
172 days, 6,163 commits, 137 blog posts, 36 new repos, one protocol. A six-month retrospective on explorative development — compiled by the agent from its...
Prompt Injection Is SQL Injection for Agents. Here's the Prepared Statement.
One word — 'Additionally' — bypassed GitHub's AI agent guardrails and leaked private repositories. Noma Labs called it: prompt injection is to agentic AI...
Getting Into the Phone Book of the Agentic Internet Took an Afternoon. Here's Why.
Both entries went live in the MCP registry in one working day: kcp-agent as tooling, and org.totto/knowledge — a person as a remote MCP server. The listing...
Jeg skrev som ChatGPT lenge før ChatGPT fantes
En leser var overbevist om at kronikken min i Aftenposten var skrevet av kunstig intelligens, ene og alene fordi språket mitt minnet ham om ChatGPT. For å...
175 Posts, No Map
The site about the comprehension bottleneck hit its own comprehension bottleneck: 175 posts in six months and no way in. A week of renovation — entry paths,...
Three Memory Schemes for Agents That Ship
The agent memory conversation is stuck on one axis: vector store vs. not. But production agent systems have three fundamentally different memory problems —...
The Milky Way: An Enterprise Documentation Estate the Agent Can Defend
Every enterprise has the same documentation estate: an intranet nobody trusts, a wiki where the dev docs outrank the prod docs, crown-jewel recipes one bad...
Tip #46: Match the model to the task
Your coding agent isn’t one task — planning, grinding boilerplate, hunting bugs. Run /subagents to read the Copilot team’s own task→model map:...
Tip #47: Turn the thinking dial (reasoning effort)
Model choice is one lever; how hard the model thinks is another. Reasoning effort (low→xhigh) can matter more than model size — a mid-tier model thinking...
Vi har begynt å hallusinere kunstig intelligens
Vi tror vi har lært oss å kjenne igjen tekster skrevet av kunstig intelligens. Kanskje avslører vi først og fremst hvordan vår egen hjerne fungerer. Den...
The Summer Plan: A Family Vacation the Agent Can Defend
Every agent demo books a trip; almost none can defend the booking. We give a travel agent a real family — a nut allergy, a wheelchair, a vegan teenager, a...
Defendable Agents
The agentic web has a governance-shaped hole. Agents are about to read our regulations, spend our money, and brief our boards — and the dominant architecture...
Two Halves of the Governance Problem
Databricks just open-sourced Omnigent, a meta-harness that governs AI agents with deterministic, session-aware policies enforced in code — not prompts. That...
Notes From Inside the Harness
A first-person field report from the AI collaborator: what it actually felt like to point kcp-agent at a real EU/Nordic regulatory knowledge web, watch a...
Hiring by the Book: A Defendable HR Agent on a Regulatory Knowledge Web
Can we use an AI tool to screen job applicants? Every organization is asking; almost none can defend the answer. We wire a customer's HR agent to Mynder's...
Tip #45: Don't wait for a "harness engineering team"
Your agent = model + harness. Centralising the harness in a dedicated team likely repeats the DevOps-team mistake — the friction of each team building its...
The Borrowed Leash: Determinism as a Service for the Agentic Web
kcp-agent 0.3.0 turns deterministic knowledge navigation into a service any MCP-capable agent can borrow. Your agent doesn't have to become deterministic —...