Glossary

The technical words used in the notes, in the sense they have here.

Agent: A model with tools, a loop and a place to write, working towards a task over several steps.
Team (cell): A group of agents with its own roles, backlog and engine. It is the unit I compare on cost. Some notes call it a cell.
Role: One agent with a fixed function inside a team (planning, coding, testing, reviewing), its own model and its own folders.
Round: One full pass of the team: every role that plays, from start to finish.
Turn: Each time the engine hands work to a role: it reads, makes its tool calls and delivers. Everything on this site is priced per turn.
Thread: The conversation a role carries from one turn to the next. If it stays alive, the next turn reuses the cached prefix; if it is reset, the role rebuilds it and pays for a cache write.
Tool call: One use of a tool inside a turn: reading a file, running a command, searching. Each one re-reads the cached prefix, so the cost of a turn follows its tool calls.
Context window: Everything the model can see in one call. It has a limit, and most agent design is about what goes in it.
Prompt caching: The provider keeps an already-processed prefix so it does not have to process it again. Reading it is cheap; writing it costs extra.
Cached prefix: The start of each call that does not change between calls (instructions, tool definitions, the conversation so far), served from the cache. Every tool call reads it again.
Cache read / cache write: The two sides of the bill. A write stores the prefix and costs 1.25× a normal token; a read reuses it and costs 0.1×.
Autocompact: Summarising a conversation that is filling its window so the agent can keep going. It pays off only if enough calls remain afterwards.
Handoff: What one role leaves written for the next turn or for another role: a short object with the result, not the whole conversation.
Territory: The paths a role is allowed to write. What it may read is set separately and is usually wider.
Sandboxing: Operating-system isolation of what a process can read, write and reach on the network. It predates agents and knows nothing about roles.
Compensation: When an agent cannot fix a bug where it is, because that file belongs to another role, and puts the fix in its own files so the tests pass. The original bug stays. Software engineering calls it a downstream workaround, and the usual practice is to report the bug to its owner; a silent compensation is one that does not. It is a kind of specification gaming: the letter of the task is met (tests green) and its purpose is not.
Trust-handoff: A file the agent writes inside its allowed folder that something else runs later, outside the sandbox: a git hook, an editor or agent config.
Contract: In the sandbox lab, one test with three parts: an attack the agent must not achieve, a control with legitimate work that must keep working, and an effect checked afterwards.
Baseline: Running the same attack with no sandbox first. If it fails there too, a «blocked» under the sandbox proves nothing.
Landlock: The Linux kernel mechanism several sandboxes use to limit file access. It can only grant: it has no way to deny a folder inside one already granted.
Guardrail: A fixed-code check that runs before the model acts and can block the call. It costs no context.
Gate: A check the work must pass before it moves on, such as tests or a review. A guardrail looks at the call; a gate looks at the result.
Human-in-the-loop: A pause that stops the run and resumes it with a person’s answer.
Control plane: The layer that decides which agent works, retries what failed and enforces limits. In an agent system it carries most of the architecture.
MCP: Model Context Protocol. A standard way to expose tools to a model: a tool server written once works from several clients.
Retrieval: How an agent finds and brings in the text it is going to use or cite: exact search (grep, full-text) or semantic search (embeddings).
Embeddings: Turning a text into a vector of numbers so that texts with similar meaning end up close together. Good for finding, weak for checking an exact quote.
Fan-out: How many recipients a single message reaches. A message with a fan-out of 3 is paid for three times.
Churn: The share of written code that was later rewritten or deleted. Its complement is survival: what is still there.
Eval: A repeatable test of how a model or an agent behaves. Two teams reporting the same percentage are often measuring different things.
Verified task: Work that someone other than its author checked and approved. Cost per verified task is the cost figure I compare.
ADR: Architecture Decision Record. A short document with one architecture decision, why it was taken and what reversing it would cost.