A field dictionary for working with agents
The words you keep meeting when working with AI agents, said plainly — each links to a piece that goes deeper.
- agent / AI agent
- A program that doesn't just answer but acts: it reads a goal, plans, calls tools, and checks results in a loop. Read deeper →
- context / context window
- The working memory the model sees in one turn — instructions, data, history. It's finite, so what you put in is a decision. Read deeper →
- forcing function
- A structural constraint that makes the right behavior unavoidable rather than remembered — e.g. forcing the agent to prove a change before calling it done. Read deeper →
- orchestration
- Coordinating several agents or sub-steps on one larger job, then reconciling the results. Read deeper →
- blast radius
- The set of things a change can break. Read it before editing so nothing snaps where you didn't expect. Read deeper →
- gate / control gate
- A checkpoint mid-run where the agent pauses for a human to approve before an irreversible step. Read deeper →
- checkpoint / save-point
- A safe state to fall back to if a later step goes wrong — like a save-point in a game. Read deeper →
- eval / evaluation
- Measuring output quality against a standard, so you know whether a change helped or hurt — not by feel. Read deeper →
- autonomy
- How far an agent runs without you stepping in. How far to loosen it is a trade between speed and risk. Read deeper →
- prompt injection
- When untrusted data (a web page, email, file) sneaks instructions to the agent, hijacking your directions. Read deeper →
- token / token budget
- The unit of text the model processes — and the cost of context. More tokens isn't always better. Read deeper →
- silent breakage
- A failure that still produces plausible-looking but wrong output — no error thrown, so it slips through unchecked. Read deeper →
- CLAUDE.md / rulebook
- An always-loaded instruction file holding conventions and forcing functions so the agent does the right thing by default — not a README. Read deeper →
- confidently wrong
- The agent gives a wrong answer in a confident tone. Model confidence is not evidence of correctness. Read deeper →
- script the deterministic parts
- For fixed-pattern work, use code, not the AI — cheaper, surer, and repeatable. Read deeper →
- operator
- The person directing the agent. Operator skill — observing, delegating, gating — changes over time. Read deeper →
- handoff
- Passing context to a fresh session or agent when the old one fills up or ends, so the thread isn't lost. Read deeper →
Frequently asked
What is an AI agent?
A program that doesn't just answer but acts: it reads a goal, plans, calls tools, and checks results in a loop — rather than producing one reply and stopping.
How is an AI agent different from a chatbot?
A chatbot replies to a question; an agent takes actions across multiple steps and tools to finish a goal, not just produce one reply.
Do I need to know how to code to use an agent?
No. Many of the highest-value uses are repetitive office tasks — drafting, summarizing, reconciling, sorting — that need no code at all.
Where should I start?
Pick one task you repeat every week with a clear shape, and hand that one over first. A small sure win beats a big vague project.
How do I trust what an agent produces?
Don't trust the word 'done'. Make it run the real path and show observable evidence — a value, a screen, a result — before counting it finished.