AI Dictionary

Plain English

A chatbot answers. An agent can decide what to do next, use tools, check results, and keep working toward an outcome.

Example

A coding agent reads an issue, edits files, runs tests, and opens a pull request.

Why it matters

It is the umbrella term behind most new AI product claims, but the actual autonomy can range from a scripted workflow to a long-running worker.

Plain English

The product is no longer only generating content; it is participating in a workflow.

Example

A support agent classifies a ticket, fetches account data, drafts a refund, and asks a human to approve the action.

Why it matters

Agentic systems change product, security, and UX design because the software can affect external systems.

Plain English

The model is the brain, but the harness is the workbench, clipboard, tool belt, and supervisor.

Example

A coding harness gives an agent shell access, file editing, browser QA, progress logs, and rules for ending each session cleanly.

Why it matters

Many agent improvements now come from better harness design rather than only stronger models.

Plain English

It is the heartbeat of an agent: decide, act, look at what happened, repeat.

Example

Search the docs, read a result, update the plan, run a command, inspect the error, and try a fix.

Why it matters

Loops make AI useful for open-ended work, but they also create compounding cost and error risk.

Plain English

It is work that cannot be solved in one neat prompt, such as migrating an app or researching a market.

Example

An agent spends hours modernizing a codebase, running tests, fixing regressions, and documenting the result.

Why it matters

Long-horizon ability is becoming a major benchmark for whether agents can handle valuable real-world work.

Plain English

It can leave notes for its future self, resume after interruption, and avoid starting from scratch.

Example

A research agent checks sources every morning, updates findings, and keeps a running evidence log.

Why it matters

The next useful agents are less like chat tabs and more like durable workers with memory and accountability.

Plain English

Instead of one general agent, several specialists work in parallel or hand work between themselves.

Example

One agent researches vendors, another estimates cost, and a third audits the final recommendation.

Why it matters

Swarms promise speed and specialization, but they need orchestration, shared state, and safety boundaries.

Plain English

Instead of asking once, you pin the outcome the agent should keep pursuing until it finishes or hits a stop rule.

Example

A developer starts `/goal reduce bundle size by 20% and verify with build output`, then lets the coding agent continue until it can prove the result.

Why it matters

Goal-style commands turn agent work from a chat request into a bounded task contract with acceptance criteria.

Plain English

The agent gets a reliable save system for long work, not just a long chat history.

Example

A procurement agent pauses for budget approval, stores its state, and resumes from the same task after approval arrives days later.

Why it matters

Durability is becoming essential for production agents that run beyond a single request-response cycle.

Plain English

The agent stops at the important moment, asks for the missing decision, then continues with the answer.

Example

A refund agent pauses before issuing credit and asks a manager to approve the amount.

Why it matters

Interrupts make human guidance possible during the work, not only after the agent has already acted.

Plain English

A team of narrow agents can be easier to steer than one giant all-purpose agent.

Example

A planning agent delegates to a browser agent, a coding agent, and a QA agent.

Why it matters

Multi-agent patterns are becoming common when context grows too large or expertise needs to be separated.

Plain English

It is the project manager of a multi-agent setup.

Example

A supervisor decides whether the legal, billing, or technical sub-agent should handle a customer question.

Why it matters

Supervisor patterns make complex systems debuggable by keeping a clear owner for routing and synthesis.

Plain English

A specialist you call when the main agent should not carry every detail in its own context.

Example

A QA sub-agent opens the app in a browser and reports visual or functional regressions.

Why it matters

Sub-agents reduce context clutter and make ownership clearer in larger agent systems.

Plain English

The main agent can ask a specialist for one job and keep ownership of the final answer.

Example

A travel planner calls a visa-checking agent as a tool, receives the result, and continues planning the trip.

Why it matters

It offers specialization without losing the thread of who owns the user-facing response.

Plain English

It is less one helper and more a small crew of AI workers, each with its own workspace.

Example

A product agent drafts requirements while a coding agent implements and a review agent checks risks.

Why it matters

Agent-team vocabulary is replacing vague swarm language when systems need clearer roles and accountability.

Plain English

The AI equivalent of routing a case from support to billing, with context carried along.

Example

A triage agent hands a refund request to a billing specialist after collecting the order ID.

Why it matters

Good handoffs prevent lost context, repeated questions, and specialists acting without the right permissions.

Plain English

One agent figures out the plan; other agents do the pieces.

Example

A codebase migration agent identifies affected packages and assigns each package to a worker.

Why it matters

This pattern handles tasks where the steps are not known ahead of time.

Plain English

One AI writes; another AI reviews; the first AI revises.

Example

A drafting agent writes a policy summary, then an evaluator checks completeness and flags unsupported claims.

Why it matters

It can raise quality when the success criteria are clear enough to evaluate repeatedly.

Plain English

The AI can prepare the action, but a human still signs off.

Example

An agent drafts a customer refund, but a manager approves it before money moves.

Why it matters

It is one of the simplest ways to limit harm while still getting automation benefits.

Plain English

The system can act, but people can see what it is doing and take the wheel.

Example

A security operations agent remediates low-risk alerts automatically and escalates unusual cases.

Why it matters

As agents run for longer, oversight has to be continuous rather than one approval at the end.

Plain English

Not every agent deserves the same keys.

Example

Level 1 drafts recommendations, level 2 acts after approval, and level 3 acts within a budgeted sandbox.

Why it matters

Clear autonomy levels help teams align product promises, permissions, risk, and user trust.

Plain English

It is the model's short-term workspace.

Example

A long-context model can inspect a whole repository or a large batch of documents before answering.

Why it matters

Large context helps, but long tasks still need retrieval, memory, and summarization discipline.

Plain English

Prompt engineering asks the model well. Context engineering gives it the right room to think in.

Example

A coding tool includes the active file, related tests, recent errors, project rules, and only the relevant docs.

Why it matters

For agents, context quality often matters more than clever wording.

Plain English

Memory is what lets the system remember beyond one chat or one run.

Example

An agent stores that the user prefers concise summaries and that the project uses Bun, not npm.

Why it matters

Memory improves continuity, but it also creates new risks such as stale, sensitive, or poisoned context.