APPENDIX A: GLOSSARY OF TERMS AND ACRONYMS

TERMS BEGINNING WITH T Tamper-evident: changes would be detectable. Task intake: receiving a user's request and shaping it into a clear job. Task success metric: a measure of whether the whole job was done correctly. Taxonomy: a system for sorting things into named groups. Teaching to the test: tuning a system to score well on a known test rather than to be good generally. Telemetry: information automatically sent from a running system for study. Temperature: a setting that controls how random or creative a model's output is. Terminal: a window for typing text commands to the computer. Termination condition: a rule that decides when the agent should stop. Test: a small program that checks whether code behaves correctly. Test-driven development: writing a failing test before writing the fix. Threshold: a limit that, once crossed, triggers an action such as compaction. Throughput: how much work a system completes in a given amount of time. Ticket: a recorded request for help. Tiered storage: keeping information in levels, from fast and small to slow and large. Time to first token: wait before output begins. Timeout: a limit on how long to wait before giving up on an action. Token: a small chunk of text, often part of a word, that a model reads and writes. Token budget: a cap on how many tokens may be used. Token economics: how the number of tokens used drives the cost of running a model. Tokenizer: the program that splits text into tokens. Token-threshold trigger: a rule that starts compaction when the context reaches a set fraction of the window. Tool: a function the agent can call to affect or look at the outside world. Tool calling: the model asking the harness to run a particular tool with particular inputs. Tool definitions: the descriptions of the tools the agent can use. Tool gating: controlling which tools the model can request. Tool poisoning: hiding malicious instructions inside a tool's description. Tool registry: the catalog of available tools, with their descriptions and permissions. Tool versioning: giving each tool a version label so changes can be tracked and old ones retired safely. Total cost of ownership: the full cost of a choice, including setup and upkeep. Total operating cost: everything required to run the agent, not only token charges. Trace: the full chain of spans for one user request. Tracing: recording each step of an agent's work so the whole chain can be reviewed. Training: the process of tuning the parameters by guessing hidden text and correcting errors. Transparency: openness with users that they are dealing with automation. Transport: the method by which messages travel between a client and a server. Triage: sorting by seriousness. Trust boundary: the line between material you trust and material you must treat as possibly hostile. Type checking: confirming values are used in the ways the code declares. Typed handler: a category base with shared checks.