Blog
Notes on coding agents, traceability, and delivery, written while we build DETENT.
How to Install Claude Code, and What to Configure Before Using It on Client Code
How to install Claude Code on macOS, Linux and Windows, how to check it works, and what to set up before pointing it at a client's repository: permissions, limits and logs. Current as of 7 October 2026.
ReadSpec Kit Writes the Spec. It Does Not Prove the Code Matches It.
GitHub Spec Kit standardises specification, plan and tasks before an agent writes code. What it covers well, and what it leaves unproven once the diff lands.
ReadPrompt Engineering Is No Longer the Bottleneck. Verification Is.
Prompt engineering is the craft of phrasing instructions for a model. With coding agents the breaking point has moved downstream: to how work is verified and what is left once the session ends.
ReadEnterprise AI Agents Are Being Bought. Who Signs What They Ship?
Companies buy enterprise AI agents by function, but few contracts say who answers for what the agent delivers. A way to draw the line once: automatic, assisted, human-only.
ReadProvider or Deployer Under the EU AI Act? The Role Question for Software Built for Others
Under the EU AI Act your role depends on what you deliver and under whose name, not on the tool you used to write the code. How to tell which box you are in, using the Regulation's own definitions.
ReadShadow AI Is Not Just Data Leaving. It Is Code Arriving.
Every shadow AI report tracks employees leaking data to unapproved tools. None of them track the opposite risk: AI-written code merging into the product with no record of where it came from.
ReadClaude Code: What It Does, How to Use It, and What It Leaves Behind
A full walkthrough of Claude Code: how to install it, what happens inside a session, and what actually stays on disk, in plain text, for 30 days, once you're done.
ReadAI-Assisted Software Development Is the Destination, Not a Stage
Assisted, augmented, autonomous: what the words actually mean, why the middle one is where most contracted software work ends up, and why the signature stays human on purpose.
ReadAI Agents on Contracted Software Work: What Changes When the Code Belongs to a Client
What changes for a software house doing contracted work when an AI agent writes the code: who owns it, who answers for it, and what has to ship alongside it.
ReadAgentic Development, Defined Without the Hype
Agentic development is the short definition a non-technical buyer needs: what makes a process agentic, what it is not, and what changes for a team once it applies.
ReadWhat CI/CD Is, and What Changes When an Agent Proposes the Change
CI/CD isn't just automated builds and tests. Once an agent is the one proposing changes, the pipeline becomes the one place in the flow where proof of what was actually checked can land.
ReadAI Software Engineering: What the Discipline Looks Like When Typing Is Free
AI software engineering names the whole discipline, not one engineer's day. The practices built to manage the cost of writing code and the ones built to manage its risk are recalibrating in opposite directions.
ReadWhat Claude Code Actually Costs a Team (and What the Price List Leaves Out)
Claude Code's price depends on the Claude plan you pick, not a separate line item. What changes across Pro, Max, and Team, what happened to usage limits in September 2026, and the cost no pricing table shows.
ReadAgentic Engineering Is a Job, Not a Tool Choice
Agentic engineering is what a software engineer's day looks like once a coding agent does the typing: the skills that moved to the centre, and what still has to stay human.
ReadThe EU AI Act: What Changes for a Software House, Not a Model Trainer
The EU AI Act regulates AI systems, not code written with an AI assistant. What actually changes for a software house, and what a client asks anyway.
ReadThe Agentic SDLC: Which Stages Actually Change, Which Just Get Louder
Agentic SDLC is a category vendors are naming right now. Almost every definition says agents do more at each stage. Almost none says what has to survive each one.
ReadNIS2 and the Software Vendor: What the EU Directive Asks of Your Client's Suppliers
NIS2, the EU cybersecurity directive now in force in Italy, almost never applies to the software vendor directly. It applies to the supply chain around it, which does.
ReadAI Software Development: Why Writing the Code Was Never the Hard Part
AI software development gets judged by how much code an agent produces. That was never the bottleneck: analysis, context, verification, and acceptance are the four steps that didn't get automated, and that's where the effort moved.
ReadYour AI Governance Framework Covers Models and Data. Who Governs the Code?
AI governance frameworks answer which models are allowed and what data leaves the company. None of them answer who approved a specific agent-written change, or how you'd prove it later.
ReadDevin Desktop's Kanban board: who tracks what each agent did
Devin Desktop turned Windsurf into a Kanban board of parallel agents. The board organizes the work. It doesn't answer who can show, later, what each agent actually did.
ReadClaude Code Skills Package the Method. They Do Not Record What Happened.
Claude Code skills bundle instructions into a shareable, versioned folder. That solves how an agent gets briefed, not what a reviewer needs to check after it ran.
ReadOpenCode: The Open Source Coding Agent, and What It Leaves Unproven
OpenCode is MIT-licensed, self-hosted, and doesn't store your code by default. That also means no vendor is building the audit trail for you. What a team has to add.
ReadContext Engineering: What a Coding Agent Is Allowed to See, and What Proves It
Context engineering names what fills a coding agent's context window. Almost every guide explains how to build that pipeline. Almost none asks how a team proves, later, what was actually in it.
ReadAGENTS.md: It Tells the Agent What to Do, Not What It Did
AGENTS.md is fast becoming the standard way to instruct a coding agent. It's an instruction file, not a record of what happened. Why that distinction matters before you rely on it.
ReadClaude Code Subagents: What Changes When One Agent Delegates to Another
Claude Code subagents split a task into isolated sessions with their own context. Delegation solves a real problem inside the session. It multiplies a different one once the session ends.
ReadClaude Code Hooks: How to Turn Agent Actions Into an Audit Trail
Claude Code hooks intercept what the agent is about to do and can log it. A hook log is not a signed record of delivery on its own. What has to be added before it counts as one.
ReadClaude Agent SDK: what remains as proof when you build your own
The Claude Agent SDK ships the same engine as Claude Code. It doesn't ship the record of what your own agent did, or why. That part is on the team that builds it.
ReadVibe Coding: What Remains as Proof When the Vibe Wears Off
Vibe coding ships fast because nobody reviews every line. What a software house needs to keep from that session before the code reaches a paying client's production.
ReadClaude Code's Auto Mode Is Default. Here's What Still Needs a Signature.
Anthropic made auto mode the default in Claude Code on August 14, 2026, removing step-by-step approval prompts. What a classifier blocking bad commands does not replace: a human signing off on what shipped.
ReadWhat Is Agentic Coding? What Should Remain When the Loop Stops
Agentic coding is a loop, not a feature: plan, edit, run, verify. Most guides explain the loop and its risks. Almost none ask what should still be there once it stops.
ReadWhat an SBOM Is, and Why AI-Generated Code Needs One Too
A software bill of materials tells you what's inside a build. It doesn't tell you who, or what, decided to put it there. Why that gap matters more once agents write the code.
ReadGitHub Copilot vs Cursor: the question GitHub teams actually ask
Same two tools as Cursor vs GitHub Copilot, searched in reverse. What changes when GitHub is already the system of record before anyone opens Cursor.
ReadAI Code Review Tool: What to Buy if Review Is the Job
CodeRabbit and Greptile both catch what a human reviewer might miss on a pull request. Neither one proves a delivery was correct once the PR is merged.
ReadBest AI for Coding 2026: A Traceability Roundup
Most best-AI-for-coding lists rank the same capability scores. This one adds the column they skip: what a team can still show once the agent has finished.
ReadAI Coding Agents: What to Judge After the Agent Finishes
AI coding agents get ranked almost entirely on how well they write code inside a session. What's missing from every roundup: what a team can still show once the session is over.
ReadHow to Choose an AI Coding Assistant
Autocomplete, in-editor agent, or headless session: the criteria software houses actually use to pick an AI coding assistant, and the one most comparisons skip.
ReadHow to Evaluate Claude Code Before You Give It to a Team
Claude Code looks different once it stops being one developer's tool and becomes something a whole team depends on. What to check before rollout, and what to keep once you do.
ReadCursor vs GitHub Copilot: what a software house should ask
Cursor and GitHub Copilot both write code faster inside a session. The harder question is what a software house is left holding once the session ends.
ReadClaude Code vs Cursor: what remains as proof
Claude Code against Cursor isn't just a question of who writes code faster. What a team is left holding, as proof, once the agent has finished the work.
Read