Glossary

Reference

I have avoided jargon throughout. Where a term was unavoidable, it is here.

The basics

Agent — The thing you are building. Software that perceives, decides and acts over multiple steps, rather than only answering a single prompt.

Model — The engine inside the agent. It reasons and produces language. It is not the agent: the agent is the model plus its memory, tools, permissions and job.

Platform — The software that runs the agent, connects it to messaging services, manages memory, and schedules work. Mine is Hermes Agent.

Household context — Everything the agent knows about your family: who is who, how things are done, what matters.

Access and control

Least privilege — Giving the minimum access the job needs, and no more.

Scoped credential — A key that grants access to exactly one thing, so a leak is bounded. The opposite of one master password everywhere.

Allowlist — A list of things permitted, where everything else is refused by default. The opposite of a blocklist, which permits by default.

Approval gate — A point where the agent stops and asks a person before continuing.

Revocation — Withdrawing a credential. The measure of good access design is how cleanly this works.

Lane — One area of the household with its own agent, its own memory and its own boundaries. Mine: household, family, health, business.

Doing the work

Tool — A specific action the agent can take: create a calendar event, search the web, read a mailbox.

Scheduled job — A task that runs because a time came round, not because anyone asked.

Watcher — A scheduled job whose purpose is to notice a change and alert only when something moves.

Digest — A scheduled job that gathers several things and reports once.

Heartbeat — A regular signal that the agent is alive and working.

Self-healing — The agent restoring itself to a known state, without a person.

Memory

Working memory — The small, curated file loaded into every conversation. Mine has a hard size limit, which keeps it high-signal.

Wiki — A folder of markdown pages, one per subject, with an index and a log. Where the detail lives.

Frontmatter — The short block of metadata at the top of a markdown page: date, source, status.

Inference — Something the agent concludes rather than being told. Dangerous in memory, because a stored inference becomes an instruction nobody gave.

Consolidation — Merging recent detail into durable summaries. Runs overnight.

Provenance — Where a piece of knowledge came from. Essential for deciding whether to trust it.

Failure and safety

Silent failure — Something stopped working and nothing noticed. The failure mode most worth designing against.

Blast radius — How much damage a single mistake can do. Access design is largely the art of shrinking it.

Prompt injection — Text in something the agent reads, written to be mistaken for an instruction. The core unsolved risk in this class of system.

Rollback — Returning to an earlier known-good state. Distinct from restore, which returns to a backup.

Watchdog — An independent check that complains when the thing it watches has gone quiet.

Two words I avoid

Autonomous — I avoid it. Mine is not autonomous; it is delegated. It works within permissions I set, and stops at the things that matter. The distinction is not pedantry: “autonomous” implies judgement I have not given it.

Assistant — Used often, and fairly. But note it does the work around a decision, never the decision itself.