Newest first, grouped by year. Search or filter by topic.
Execution is not one heroic pass - it is a three-step inner loop in which the agent codes, the output is read against the specification, and over-engineering is actively stripped out. The loop is my extension to the borrowed skeleton, and this is how it runs.
The Review phase is where a human takes an agent-graded plan and does the three things no score can - raise clarity, raise completion, minimise complexity - then encodes intent before a line of code exists. Why I call its exit the workflow's most important handoff.
PREVC's Planning phase has the agent decompose the work the way project managers always have, then do something no methodology asked of it - grade its own plan. The work breakdown structure, the fifteen-minute waterfall, and the three scores I added.
Agents now work unattended for stretches long enough that the conversation can no longer save you - why unstructured runs fail, the five-phase PREVC workflow I adapted from an open-source skeleton, and where this ten-part series goes.
The honest retrospective this series owes its readers - the thesis I believe but cannot cite, the threshold that is really one team's habit, the discipline with no how-to, the figures that will age, and the work you should not hand to an agent at all.
Dan Shapiro's Five Levels ladder maps where the industry actually stands with autonomous coding agents - most of it lower than it believes - and StrongDM's three-person software factory shows what the top rung looks like when someone builds it.
The disciplines tell you what to learn - the primitives are what you actually write. Five small, recurring artefacts of AI-directed work, from the self-contained problem statement to the evaluation harness, each shown with a worked example.
The highest-leverage discipline in the stack is writing down what done looks like - what a specification must contain before an agent can execute for hours against it, and what Addy Osmani's 80% problem actually says about the work that remains.
An agent can have perfect context and a flawless harness and still optimise for the wrong thing - what Klarna's 2024 launch figures and 2025 reversal actually teach about encoding organisational intent, and three mechanisms that stop that intent drifting.
Progressive disclosure at repository scale - how a hundred-line map steers agents through a million lines of code, what OpenAI's Codex experiment actually reported, and the benchmark run where changing only the harness moved an agent from Top 30 to Top 5.
Why the context file that says what an agent must and must not do stays current for months, while the one that describes how the code works is stale by Friday - and how to decide what loads always versus what gets fetched on demand.
The permanent infrastructure that makes every agent session start smart - tool configuration, codebase onboarding and workflow automation, and the evidence that the harness moves results more than the model does.
The workflow that turns one long degrading session into three small clean ones - what each phase produces, why the artefacts matter more than the phases, and how much ceremony a task actually deserves.
Context windows degrade before they fill - the evidence for context rot, one team's 40% working rule, and the three ways a window turns toxic long before it runs out of room.
Write and select decide what enters the context window - compress and isolate decide what survives inside it, from intentional compaction to sub-agents and the file system as the agent's real memory.
LangChain's four-pillar taxonomy gives context engineering its working vocabulary - this part walks the first two pillars in depth, authoring what the agent wakes up knowing and choosing what a task actually needs.
Everything that looks like memory in an agent is tokens being written back into the window - the mental model beneath all of context engineering, and the week in June 2025 when the discipline got its name.
The prompt is the smallest and least durable artefact in the whole agent system - why the original skill of the AI era stopped being the differentiator, and what good prompt craft still buys you every single day.
Between late 2025 and early 2026 coding agents stopped pausing to ask questions - this series opener maps the five-discipline stack I now use to direct them, from table-stakes prompting to specification engineering.
A Christmas Day commit on an unfinished feature, three years of unactioned dependabot, a final fix by a stranger - and the estate's real lesson, which is about re-reading rather than writing.
A product whose entire interface is a YAML document, shipped in an image that document cannot enter - plus the resilience inventory of a gateway with retries and nothing else.
CI builds Release, where six extensions compile against NuGet instead of your working tree - and the only thing that notices is a project with no tests in it.
One commit reclaimed a clean package id and deleted the version prefix in the same breath - leaving two identities on the public feed with inverted version ordering and a csproj that cannot build the package it consumes.
A caller-identity feature added at 21:10, disabled at 22:31 and deleted four days later - and the residue is a gateway that carries no caller identity into a message at all.
Of eight JWT keys in the README's flagship example, four bind to nothing, three restate defaults and one has any effect - and the last edit to that README came ten minutes before the commit that invalidated it.
