Glossary.

The words this site keeps using, defined once. Some are ordinary terms used in a narrower sense than usual; some were coined here because nothing else named the thing. Each links to where it's argued in full.

A test suite that tells the truth: Also: the agentic prerequisite; A suite where a red build is still information. It's the variable that decides whether agents make a team faster or just louder. A team with no tests knows it's flying blind; a team with a suite it has stopped fully trusting thinks it has instruments and doesn't.; Argued in The agentic prerequisite. Also in Nobody has to approve the agent's merge, Code freezes are a confession.
Autonomy tiers: Also: autonomy levels, the autonomy dial; A system that sorts agent-authored change into a small number of classes by how expensive it is to be wrong, not by what kind of task it is, and gives each class a named gate and a named owner who can move work between classes. Write it before the first bad merge: the version written after an incident is a restriction, the version written before one is a policy. A tier ratchets one way; incidents tighten the gate rather than shrink the grant.; Argued in Set the autonomy tiers before you need them. Also in Autonomy levels and blast radius, Build an agentic-first delivery org.
Blast radius (for agents): What an agent can read, what it can write or execute, and how far a mistake can travel before any human sees it. That scope, not how hard the task looked, is what sets your exposure. Before granting autonomy over anything, ask what undoing a mistake would cost and who notices before it compounds.; Argued in Autonomy levels and blast radius. Also in Nobody has to approve the agent's merge, Review is not the bottleneck. Trust is..
Characterization tests: Also: characterisation tests; Tests that check what an old system actually does, not whether it does the right thing, including the parts nobody would defend if asked. Run the current system against a wide slice of real inputs, record the outputs, and turn them into assertions before a line of the replacement exists. Every mismatch between old and new is a gap that's cheaper to find in a comparison log than in a customer's inbox.; Argued in Migrate the legacy service with agents, seam by seam. Also in Never rewrite was a budget rule, Where agents earn their keep.
Construction and rediscovery: Also: the two bills of a rewrite; The two costs of a rewrite. Construction is building the new system. Rediscovery is recovering the years of accumulated correctness the old system encodes. "Never rewrite" worked as a rule because the first bill made the second unbearable. With agents, rewrites are cheap to build and still expensive to specify.; Argued in Never rewrite was a budget rule. Also in Migrate the legacy service with agents, seam by seam.
Cost per merged change: AI spend divided by merged changes, broken out by class of change. It's the unit that exposes a workflow producing the right answer at ten times the expected cost, a defect you'll never find while your eval only asks whether the output was right. AI spend is a design defect until proven otherwise, and the person who can fix it doesn't work in finance.; Argued in Your AI bill is a design defect. Also in Adoption isn't a result.
Intent architecture: Also: intent document; Stating what you actually want to be true, precisely enough that a system can check whether it's true without asking you, and letting the system work out how to get there. The written form is an intent document: one page naming the outcome, the trade-offs, the open questions, what would reopen it, and a named owner.; Argued in Intent architecture. Also in Intent document builder, The substitution trap.
Invoked vs ambient AI: Also: ambient mode; Invoked AI waits for you to bring the context: open a window, explain the problem, paste the code, describe the constraints, and start cold every time. Ambient AI already holds the context and surfaces what needs your judgment. The skill that matters moves from prompting well to setting intent clearly.; Argued in The substitution trap. Also in The org chart is the bottleneck.
Organisational attention: The genuinely scarce resource in an engineering org, and the denominator of high performance: shipped outcomes per unit of attention. A high-performing organisation spends very little attention per delivered outcome. A slow one spends enormous amounts of it and produces meeting notes.; Argued in What high performance actually means. Also in Adoption isn't a result, Agents are not headcount.
Process debt: Rules and gates that outlived the failure they were written to prevent. It's harder to remove than technical debt: bad code has no defenders, but a rule has an author, a founding story, and usually an incident people still remember, so removing it looks like inviting the incident back. Keep it payable by writing each rule down with the failure it prevents, when that last happened, who owns it, and when it will be reviewed.; Argued in Process debt. Also in The gatekeeping reflex.
Prompt tax: The recurring cost of operating an AI workflow, as distinct from the cost of building it. It's the price paid every single cycle, in time, attention, and organisational fragility, for as long as a human has to sit in the loop and translate. Nobody notices the total until the one person paying it goes on leave or takes a new job.; Argued in The prompt tax. Also in Prompt tax calculator, Intent architecture.
Review capacity: Also: reviewer capacity; How many diffs a person can carefully read and reason about in a day. It's roughly fixed, and it doesn't rise with how excited anyone is about tooling. When agents make producing diffs much faster and deciding whether a diff is safe no faster at all, you aren't buying speed. You're buying a longer queue with a faster front door.; Argued in Review capacity is the new bottleneck. Also in Team composition when review is the constraint, Turn your review queue back into flow.
Service as software: Consuming delivery software as a service instead of configuring and operating it. You push code; the service validates, tests and deploys, and you pay for successful deployments rather than pipeline minutes. What you stop owning is the mechanism: which jobs run, in what order, on which machines, with which caches.; Argued in Service as software.
The context is the program: Also: context files are production code; For an agent, context files such as AGENTS.md and CLAUDE.md are executed literally, so they are part of the program, not documentation about it. Degraded agent output is what a context problem looks like from the outside. Whoever changes the behaviour a context file describes updates the file in the same change.; Argued in Context is maintenance, not a build. Also in Context files are production code, and yours has no owner, An AGENTS.md that survives six months.
The four loops: Also: delivery, learning, people and strategy loops; An engineering organisation doesn't run one loop. It runs four, delivery, learning, people and strategy, and they don't share a clock. The rate at which real value reaches customers is set by the slowest loop rather than the fastest, and the delivery loop gets blamed for problems it didn't cause because it's the one everyone can see.; Argued in The four loops every org runs. Also in The org chart is the bottleneck.
The gate, not the signature: Also: approval as proxy; The safety of a merge was never the reviewer's approval. It's the machinery underneath it: a test suite that tells the truth, every change behind a flag, a staged rollout, a rollback measured in seconds, and a blast radius small enough that being wrong is cheap. Approval was a proxy, and proxies stop measuring what they used to once the volume changes. At forty diffs a day a reviewer isn't exercising judgment; she's demonstrating availability.; Argued in Nobody has to approve the agent's merge. Also in Review is not the bottleneck. Trust is., Elite is now table stakes.
The gatekeeping reflex: Answering faster output by adding a room full of people to the path and calling it governance, instead of making validation a property of the pipeline. The gates don't improve quality; they reduce anxiety. A control exists to catch something, while a gate built out of anxiety exists to be passed through.; Argued in The gatekeeping reflex. Also in Delete the AI council, Nobody has to approve the agent's merge.
The Kubernetes tax: The cost a small team pays for adopting Kubernetes before it needs it. The bill isn't the weekend it takes to set up. It's every subsequent debugging session running through a layer nobody on the team fully understands, forever. That's the real price, and it's not on the invoice.; Argued in The Kubernetes tax. Also in The monolith default, Boring technology, deliberately chosen.
The rule of two-or-three: Also: wait for the repeat; Don't build structure (a career ladder, a new process, a platform team, a hard conversation) until the failure it answers has happened a second or third time. The first occurrence tells you almost nothing about whether a problem is structural. The second tells you a great deal, and what you build then describes the actual pattern instead of the one vivid instance on everyone's mind.; Argued in Process debt. Also in Delete the AI council.
The substitution trap: Also: substitution mode; Using AI to take the work as it is, the same steps in the same order with the same people handing it to each other, and make each step a bit cheaper. The maximum upside is a fraction of current labour cost: a cost optimisation story, with the ceiling fixed the moment success was defined as hours saved on existing tasks.; Argued in The substitution trap. Also in The org chart is the bottleneck, Adoption isn't a result.
Trust is a price: Everyone in an organisation carries an estimate, updated constantly, of what it costs to say a true and unwelcome thing out loud. The leader sets that price, not with a values statement but with a few seconds of their own reaction. Safety is about what happens when you tell the truth, not about what happens when you don't do the work.; Argued in Trust is a price, not a poster. Also in Build an agentic-first delivery org.
Two planes: Also: agent plane, production plane; Separating where agents work from where production lives. The agent plane is ephemeral, synthetic-data-only, and priced on boot time. The production plane is durable, holds real data, and changes as little as the work allows. The only thing they share is the gate: an agent puts a PR on it, and that's the entire extent of its reach. If a network path runs from a sandbox to production, you have one plane and a naming convention.; Argued in Build an agentic-first delivery org.
Unowned code: Code that passed every check, does what it was asked to do, and has no person behind it who remembers why it's written that way. It isn't technical debt, because nobody knowingly borrowed anything: it's an orphan. The fix for a loan is a repayment plan; the fix for an orphan is a name. Treat an unowned merge as unmergeable, regardless of who or what authored it.; Argued in Agents don't create technical debt. They create unowned code.. Also in Ownership is the atomic unit, Diagnose an AI-forward org in a week.