Claude agents talking to each other
I started three Claude sessions in parallel in one codebase, each as a product owner, each orchestrated by Fable 5.1. While they worked, they sent each other messages: announcing merges, recording the commit SHA afterwards and coordinating the multi-hour run of the test suite. This article shows the setup, excerpts from the three terminals and the route the messages take technically.
The Limits of Human Sign-Off in Agentic Software Development
A human sign-off in an agentic SDLC certifies a fixed state of the work against criteria defined in advance. It does not cover the agent run itself. My experience and five papers from spring and summer 2026 point to cognitive overload. Placing a human in the loop can create the appearance of safety and participation.
A junior developer for a one-time €10,000?
A local 27B model produces a usable web page on a MacBook Pro. The experiment shows how much of a closely supervised junior role local models can take on and where their limits are.
Spec drift is the normal state
Spec-driven development promises that a precise specification determines the outcome of the agents' work. In two internal projects with around 850,000 lines of product code between them, more than 5,000 ticket files and more than 7,000 commits, specification and code nevertheless kept diverging. This post backs that with primary data from both projects and ten recent studies. It also describes the countermeasures from practice, their cost, and why all of it remains mitigation.
The second lane: why agents duplicate code in existing systems
Coding agents in existing systems often place a second implementation next to the existing path. This article explains the phenomenon, backs it with my own analysis of correction findings from a branch's Git history, and describes a review process with a planning gate, an independent auditor, and a machine-readable verdict.
The Development Role: When Nobody Types Code Anymore
In the agentic engineering pipeline for the SDLC, the developer no longer writes source code line by line. Her work shifts to feasibility, architecture, ticket slicing, plan, and review. This post in the series shows why developer and architect merge, how the agent takes over the role, and why one hundred percent test coverage suddenly becomes affordable without guaranteeing correctness.
The Business Analysis Role: What Should Be Built, and Why
The second part of the series on the agentic engineering pipeline for the SDLC turns to business analysis. It delivers a documented truth: what should happen, for whom, on what factual basis and with which open questions. The post shows why this task is not a new invention, how agents make the old goal of a complete, testable ticket reachable, and which business E2E contents must exist before the handover to development.
The Agentic Engineering Pipeline for the SDLC
Agentic tools speed the work up. In brownfield they often produce errors that surface only late. This opening piece of a series introduces the agentic engineering pipeline for the SDLC: fixed roles, exploratory phases, state in versioned artifacts, a quality gate at every handover, and a second model opinion for important decisions.
AGILE, as we know it, is disappearing
The current state of research on agentic software development: agile practices can no longer be adapted, because their economic, methodological and personnel assumptions are falling away.
The Architecture of Forgetting
Language models have no memory. What looks like recollection is a reconstruction at every call. A comparative analysis of how six agentic CLIs solve the problem of a scarce working memory architecturally, with precise thresholds and verifiable citations from the respective codebases.
The ROI Depends on the Foundation: What Google's DORA Team Says About AI-Assisted Development
In February 2026, Google Cloud published a 60-page report on the return on investment of AI-assisted software development. In it, the DORA team and Google’s consulting unit “delta” model the first-year costs of tool adoption, learning curves, and value creation. The central lesson lies outside the tables: AI is an amplifier. Without a solid foundation, it amplifies an organization’s dysfunctions.
Agentic Coding Forces Development Teams to Adapt
Language models have left the editor and moved to the command line. They plan, write, test and deploy. Anyone working without a method loses control over the result. I have developed a training programme that outlines what a team needs to set up end-to-end to remain competitive.
Where the agent touches the wall
My experimental project AEGIS checks the actions of autonomous AI agents against a formal set of rules before they are executed.
The Algorithm: How DDIC Resolves Conflicts
Two rules contradict each other: a nurse is not allowed to read patient records, but in an emergency, she is. Which one wins? How does the engine prove this in microseconds, without a theorem prover and without magic?
When everyone suddenly wants AI
When companies want to roll out AI coding tools at scale, it rarely fails because of the model, more often because of context, habits and weak processes.
When AI Writes the Rules
How do we get from 'doctors are allowed to view patient records' to a formal rule that a computer can verify deterministically? On rule languages, translation gaps, and the question of whether an LLM should be allowed to help.
The Deterministic Gatekeeper for an LLM
Why no amount of training in the world is enough to turn a language model into a rule-based system. And what the alternative is.
Who Orchestrates the Specialists?
In 1997, I posted a link to the Cyc project on my company website. Nearly 30 years later, the same question arises, only differently. About the void at the center of AI architecture.
Context Is Everything
AI agents are only as good as the context you give them. Anyone running large projects with agents gets further with better context management than with a better model. A field report from two years of practice.
Elegant Garbage
LLMs give people without expertise the tools to produce impressive-looking output. The problem lies less in the technology than in the intellectual self-deception it makes easier.
Who Checks the Last 10%?
AI agents write code faster than ever before. The ability to evaluate that code, though, cannot be automated, and it is in danger of disappearing.
With AI, many projects are 90% complete
Why AI won't replace senior developers
A Return to the Waterfall Model?
LLM agents can write code in minutes. But who determines the architecture? Why greenfield development with AI requires more planning than before.