---
title: Concept trainer
description: Spaced repetition, so what you learn stays learned. Try to say the answer out loud first, then flip the card. Got it moves a card to a longer interval (1, 3, 7, then 21 days). Again sends it back to the start and re-queues it this session.
author: Asem Abdo
course: Agentic Engineering
lesson: Flashcards
url: https://openplanr.dev/learn/agentic-engineering/flashcards
updated: 2026-10-17
related:
  - https://openplanr.dev/learn/agentic-engineering/glossary.md
---

# Concept trainer

By [Asem Abdo](https://www.linkedin.com/in/asemabdo/) · Updated 2026-10-17

*Memorize*

Spaced repetition, so what you learn stays learned. Try to say the answer out loud first, then flip the card. Got it moves a card to a longer interval (1, 3, 7, then 21 days). Again sends it back to the start and re-queues it this session.

## Agent

Foundations, level 1.

- **Definition:** A model that decides its own next step and calls tools in a loop until a goal or a budget stops it.
- **Analogy:** A contractor who chooses which tool to pick up next.
- **Say it:** An agent is an LLM in a loop with tools and a stopping condition.

## Workflow

Foundations, level 1.

- **Definition:** Fixed steps defined in code; the model fills each step but never chooses the path.
- **Analogy:** An assembly line with a skilled worker at each station.
- **Say it:** I default to workflows and add autonomy only where the path can't be predicted.

## Agent loop

Foundations, level 1.

- **Definition:** Read context, propose a tool call, harness runs it, result appended, repeat.
- **Analogy:** Think, act, look, repeat.
- **Say it:** Most bugs I fix are in the loop, not the prompt.

## Tool

Foundations, level 1.

- **Definition:** A function the model may ask the harness to run, described by name, description and input schema.
- **Analogy:** An API endpoint whose client is a model.
- **Say it:** Tool design is API design for a reader that takes every word literally.

## Tool call

Foundations, level 1.

- **Definition:** A structured request from the model naming a tool and its arguments. The harness decides whether to run it.
- **Analogy:** A work order, not the work itself.
- **Say it:** The model proposes; the harness disposes.

## Context window

Foundations, level 1.

- **Definition:** Everything the model can see in one call, measured in tokens.
- **Analogy:** A desk: only what is on it can be used.
- **Say it:** I treat the context window as a budget, not a dump.

## Tokens

Foundations, level 1.

- **Definition:** The units of text a model reads and writes; they drive cost and latency.
- **Analogy:** Metered words.
- **Say it:** I track input, output and cached tokens per run.

## System prompt

Foundations, level 1.

- **Definition:** Standing instructions that frame every turn of a session.
- **Analogy:** A job description handed over on day one.
- **Say it:** The system prompt sets role and rules; task detail arrives just in time.

## Stop condition

Foundations, level 1.

- **Definition:** The rule that ends the loop: a final answer, a passed check, or a budget.
- **Analogy:** The finish line.
- **Say it:** Every agent needs a stop condition the harness controls.

## Budget

Foundations, level 1.

- **Definition:** A hard cap on steps, tokens, time or cost per run.
- **Analogy:** A spending limit on a company card.
- **Say it:** Every autonomous run has a hard budget.

## Structured output

Foundations, level 2.

- **Definition:** Responses forced into a JSON schema so programs can consume them safely.
- **Analogy:** A form instead of a letter.
- **Say it:** Anything another program consumes is schema-validated.

## Errors as observations

Foundations, level 2.

- **Definition:** Tool failures returned to the model as results instead of crashing the run.
- **Analogy:** Telling a colleague what went wrong instead of firing them.
- **Say it:** Tool errors go back to the model so it can correct itself.

## Extended thinking

Foundations, level 2.

- **Definition:** The model spends extra tokens reasoning before it answers or acts.
- **Analogy:** Thinking before speaking, billed by the minute.
- **Say it:** I raise thinking effort for planning and debugging, not for routine edits.

## Model tiering

Foundations, level 2.

- **Definition:** Routing each step to the cheapest model that does it well.
- **Analogy:** Senior engineers design, juniors implement.
- **Say it:** Strong models plan and review; cheaper models execute.

## Human in the loop

Foundations, level 2.

- **Definition:** A person approves at defined checkpoints where mistakes are expensive.
- **Analogy:** A second signature on a large payment.
- **Say it:** I put review gates where a mistake is expensive or irreversible.

## Agentic engineering

Foundations, level 2.

- **Definition:** Building reliable systems where agents do real work: context, tools, guardrails, evals.
- **Analogy:** Engineering the kitchen, not just hiring the chef.
- **Say it:** It's software engineering where one component is non-deterministic, so verification matters most.

## Non-determinism

Foundations, level 2.

- **Definition:** The same input can produce different outputs.
- **Analogy:** Asking two colleagues the same question.
- **Say it:** I design for variance: validate, retry, never assume.

## Hallucination

Foundations, level 2.

- **Definition:** Confident output not supported by facts or sources.
- **Analogy:** A fluent guess.
- **Say it:** Grounding and citations reduce it; evals measure it.

## Context rot

Foundations, level 3.

- **Definition:** As tokens accumulate, the model gets worse at recalling and using what is in its window.
- **Analogy:** A desk so cluttered you can't find the right paper.
- **Say it:** A bigger window is not better attention; I keep context lean and compact early.

## Tool sprawl

Foundations, level 3.

- **Definition:** Too many tools: each definition costs context every call and choice gets harder.
- **Analogy:** A toolbox so full you grab the wrong wrench.
- **Say it:** Fewer, sharper tools beat a big toolbox.

## Harness

Context and harness, level 1.

- **Definition:** Everything around the model: prompt, tools, permissions, hooks, memory, context management, orchestration.
- **Analogy:** The car around the engine.
- **Say it:** I spend more time on the harness than on prompts, because that's where reliability comes from.

## Harness engineering

Context and harness, level 1.

- **Definition:** Designing the guides and sensors around an agent so its output can be trusted with less supervision.
- **Analogy:** Tuning the car, not swapping the engine.
- **Say it:** The same model performs very differently depending on the harness.

## Context engineering

Context and harness, level 1.

- **Definition:** Curating and maintaining the optimal set of tokens in the window at each step (Anthropic's definition).
- **Analogy:** Deciding what goes on the desk and when.
- **Say it:** Prompt engineering is the words; context engineering is what is present at all.

## Just-in-time context

Context and harness, level 2.

- **Definition:** Loading rules or docs at the moment they matter instead of up front.
- **Analogy:** A checklist handed over at the start of the task, not at hiring.
- **Say it:** Rules injected at the right step beat a long file the model skims once.

## Progressive disclosure

Context and harness, level 2.

- **Definition:** A short summary always loaded; detail loaded only when needed.
- **Analogy:** A table of contents with chapters on request.
- **Say it:** Skills work this way: description loaded, body on demand.

## Compaction

Context and harness, level 2.

- **Definition:** Summarizing older conversation to free context. Lossy.
- **Analogy:** Replacing meeting minutes with a one-line recap.
- **Say it:** Anything critical lives in files, not only in chat history.

## Memory

Context and harness, level 2.

- **Definition:** Facts saved to files and reloaded across sessions.
- **Analogy:** A notebook the agent reads every morning.
- **Say it:** Durable decisions go into versioned files the agent reads.

## Prompt caching

Context and harness, level 2.

- **Definition:** The provider reuses an unchanged prompt prefix so repeat calls are cheaper and faster.
- **Analogy:** Reusing yesterday's prep work.
- **Say it:** I keep stable content first so it stays cacheable.

## CLAUDE.md

Context and harness, level 1.

- **Definition:** Project instructions loaded at session start (AGENTS.md is the cross-tool equivalent).
- **Analogy:** The README written for the agent.
- **Say it:** Short and factual. If a rule must never break, it goes in a hook.

## Skill

Context and harness, level 1.

- **Definition:** A folder with SKILL.md and scripts; the description stays loaded, the body loads on demand.
- **Analogy:** A playbook pulled off the shelf when needed.
- **Say it:** Skills keep the base context small; the agent pulls the playbook when the task matches.

## Subagent

Context and harness, level 1.

- **Definition:** A separate agent with its own clean context, prompt and tools, started by a parent with a brief.
- **Analogy:** A contractor with a written brief and no history.
- **Say it:** The main agent writes the brief and reviews; the subagent does the noisy work.

## Hook

Context and harness, level 1.

- **Definition:** A command that runs on a lifecycle event and can block, inject context or log.
- **Analogy:** Middleware for an agent.
- **Say it:** Hooks turn a 'please don't' into a hard block.

## Hook exit code 2

Context and harness, level 1.

- **Definition:** Blocks the action; for PreToolUse, stderr goes back to Claude as the reason. Exit 0 proceeds; any other code is non-blocking.
- **Analogy:** A red light versus a warning sign.
- **Say it:** Guardrails exit 2, never 1.

```python
def hook(path):
    return 2 if '/migrations/' in path else 0
```

## PreToolUse

Context and harness, level 2.

- **Definition:** Hook event before a tool runs; can block, allow or modify the call.
- **Analogy:** A bouncer at the door.
- **Say it:** My migration guard is a PreToolUse hook on Edit and Write.

## PostToolUse

Context and harness, level 2.

- **Definition:** Hook event after a tool runs; used to format, lint, log or feed back problems.
- **Analogy:** Quality control after each step.
- **Say it:** PostToolUse runs the formatter so I never review style.

## Stop hook

Context and harness, level 2.

- **Definition:** Runs when the agent wants to finish; exit 2 forces it to continue. Check stop_hook_active.
- **Analogy:** A manager saying 'not done until tests pass'.
- **Say it:** My Stop hook makes the agent run tests before it may finish.

## SessionStart hook

Context and harness, level 2.

- **Definition:** Runs at session start, resume or after compaction; can inject context or set environment.
- **Analogy:** A morning briefing.
- **Say it:** SessionStart injects a one-line pointer, not a wall of text.

## Slash command

Context and harness, level 1.

- **Definition:** A saved prompt you trigger by name.
- **Analogy:** A keyboard shortcut for a workflow.
- **Say it:** Commands are for things I start; skills are for things the agent should recognize.

## Plugin

Context and harness, level 2.

- **Definition:** A bundle of skills, hooks, commands, agents and MCP servers.
- **Analogy:** A dotfiles repo you can install.
- **Say it:** A plugin is how a team ships its harness.

## MCP

Context and harness, level 1.

- **Definition:** Model Context Protocol: an open standard exposing tools, resources and prompts to any compatible agent.
- **Analogy:** USB for agent tools.
- **Say it:** MCP standardizes the integration layer so one server works across clients.

## CLI vs MCP

Context and harness, level 2.

- **Definition:** CLI when a shell and a known tool exist; MCP for no-shell clients, managed auth and SaaS.
- **Analogy:** A screwdriver you own versus a rental service.
- **Say it:** CLI is cheap on context; MCP wins when there's no shell or I need managed auth.

## Headless mode

Context and harness, level 2.

- **Definition:** Running the agent non-interactively, for example claude -p, in CI or scripts.
- **Analogy:** Autopilot.
- **Say it:** In CI the agent runs headless with tight permissions and structured output.

## Git worktree

Context and harness, level 2.

- **Definition:** Several checkouts of one repo side by side, one branch each.
- **Analogy:** Separate desks for parallel projects.
- **Say it:** One worktree per agent, so parallel work never collides.

## Plan mode

Context and harness, level 1.

- **Definition:** The agent proposes a plan before editing anything.
- **Analogy:** An architect's sketch before construction.
- **Say it:** For anything non-trivial I review the plan before any file changes.

## Permission modes

Context and harness, level 2.

- **Definition:** What the agent may do without asking: read only, edit, run commands, bypass.
- **Analogy:** Access badges.
- **Say it:** I widen permissions only inside a sandbox or worktree.

## Statusline

Context and harness, level 1.

- **Definition:** A command that renders model, context %, cost and session id in Claude Code.
- **Analogy:** A dashboard in the car.
- **Say it:** If I can't see cost and context per session, I can't improve the workflow.

## Brief

Context and harness, level 3.

- **Definition:** The prompt a parent writes for a subagent; everything the subagent will know.
- **Analogy:** A handover note.
- **Say it:** Most subagent failures are brief failures.

## tmux

Context and harness, level 3.

- **Definition:** Terminal multiplexer: windows and panes that keep running when you close the terminal.
- **Analogy:** A workspace that survives closing the laptop.
- **Say it:** One tmux window per project, agents keep running when I step away.

## Orchestrator-worker

Loops and proof, level 1.

- **Definition:** A lead agent splits the task, workers execute, the lead merges.
- **Analogy:** A tech lead and a squad.
- **Say it:** Worker quality equals brief quality.

## Fan-out / fan-in

Loops and proof, level 1.

- **Definition:** Parallel workers on independent subtasks, results combined.
- **Analogy:** Splitting a book among translators.
- **Say it:** Fan-out cuts latency and multiplies cost, so I cap it.

## Routing

Loops and proof, level 2.

- **Definition:** A classifier sends each request to the right specialist or model tier.
- **Analogy:** A hospital triage desk.
- **Say it:** Routing is where tiering saves most of the money.

## Evaluator-optimizer

Loops and proof, level 2.

- **Definition:** One agent produces, another critiques, repeat until good enough.
- **Analogy:** Writer and editor.
- **Say it:** It only works if the critic has a different view than the writer.

## Verifier / skeptic

Loops and proof, level 2.

- **Definition:** A final agent re-checks claims against independent evidence.
- **Analogy:** An auditor with their own books.
- **Say it:** A skeptic that reads the same sources just agrees.

## Delegation depth limit

Loops and proof, level 2.

- **Definition:** A cap on how many levels agents may delegate.
- **Analogy:** A rule against infinite sub-contracting.
- **Say it:** Every delegation carries a depth counter.

## Evals

Loops and proof, level 1.

- **Definition:** Repeatable tests that score agent output over a set of cases.
- **Analogy:** Unit tests for fuzzy output.
- **Say it:** I don't change a prompt or harness without an eval to compare.

## Golden set

Loops and proof, level 1.

- **Definition:** A fixed set of real inputs with known good outputs.
- **Analogy:** An answer key.
- **Say it:** I start with 20 to 50 real cases, not synthetic ones.

## LLM-as-judge

Loops and proof, level 2.

- **Definition:** A model grades outputs against a rubric.
- **Analogy:** A teaching assistant grading with a rubric.
- **Say it:** Useful at scale; I spot-check the judge against human labels.

## Regression eval

Loops and proof, level 2.

- **Definition:** The eval suite reruns after every prompt or harness change.
- **Analogy:** CI for behaviour.
- **Say it:** Every harness change runs the eval suite, like unit tests.

## Tracing

Loops and proof, level 1.

- **Definition:** Recording every step of a run: tool calls, tokens, latency, cost.
- **Analogy:** A flight recorder.
- **Say it:** With traces I can see exactly where a run got stuck.

## Cost per accepted task

Loops and proof, level 2.

- **Definition:** Total spend divided by runs whose output was accepted.
- **Analogy:** Cost per sale, not per ad.
- **Say it:** The number that matters is cost per accepted task, not per token.

## Prompt injection

Loops and proof, level 1.

- **Definition:** Untrusted content that tries to act as instructions.
- **Analogy:** A forged memo slipped into the inbox.
- **Say it:** Anything the agent reads from tools is data, never instructions.

## Lethal trifecta

Loops and proof, level 2.

- **Definition:** Private data plus untrusted input plus a way to send data out, in one agent.
- **Analogy:** Keys, a stranger, and an open door.
- **Say it:** I never give one agent all three without a human gate.

## Least privilege

Loops and proof, level 2.

- **Definition:** Agents get only the tools and access the task needs.
- **Analogy:** A visitor badge, not a master key.
- **Say it:** Read-only analyzers by default; write access is earned.

## Sandboxing

Loops and proof, level 2.

- **Definition:** Running agents in isolated environments: containers, VMs, worktrees.
- **Analogy:** A test kitchen.
- **Say it:** Bypass-permissions mode only inside a sandbox.

## Supply chain risk

Loops and proof, level 3.

- **Definition:** Third-party plugins, skills and MCP servers run with your access.
- **Analogy:** Hiring a contractor without references.
- **Say it:** I review a plugin like I'd review an npm package.

## RAG

Loops and proof, level 1.

- **Definition:** Retrieve relevant documents, add them to the prompt, then generate.
- **Analogy:** An open-book exam.
- **Say it:** RAG grounds answers in our data and makes them citable.

## Hybrid search

Loops and proof, level 2.

- **Definition:** Combining keyword and vector search, usually followed by reranking.
- **Analogy:** Searching by exact words and by meaning.
- **Say it:** Hybrid retrieval plus a reranker is my default for production RAG.

## Agentic search

Loops and proof, level 2.

- **Definition:** The agent searches iteratively with tools like grep instead of one retrieval step.
- **Analogy:** A detective following leads.
- **Say it:** For code, grep and file reads often beat a vector index.

## A2A

Loops and proof, level 3.

- **Definition:** A protocol for agents from different systems to delegate to each other.
- **Analogy:** Agents exchanging business cards.
- **Say it:** MCP connects agents to tools; A2A connects agents to agents.

## Streamable HTTP

Loops and proof, level 3.

- **Definition:** MCP's standard remote transport; it replaced the deprecated HTTP+SSE transport.
- **Analogy:** One door instead of two.
- **Say it:** Local MCP servers use stdio; remote ones use Streamable HTTP.

## Point, example, trade-off

Loops and proof, level 1.

- **Definition:** The three-part shape for every spoken answer.
- **Analogy:** Headline, evidence, fine print.
- **Say it:** Short answer first, one real example, then what it costs.

## Enforce, don't request

Loops and proof, level 1.

- **Definition:** Rules that matter live in hooks, schemas or permissions, not in prompts.
- **Analogy:** A locked door versus a 'please knock' sign.
- **Say it:** If a rule must hold, I enforce it in code.

## Harness beats model

Loops and proof, level 1.

- **Definition:** Most quality gains come from context, tools and verification, not from switching models.
- **Analogy:** A better car beats a slightly bigger engine.
- **Say it:** Before I switch models, I fix the harness.

## Vibe coding

Foundations, level 1.

- **Definition:** Accepting what the model writes without really reading it; Karpathy's term from February 2025.
- **Analogy:** Driving by feel on an empty road.
- **Say it:** Vibe coding is fine for prototypes; production needs checks.

## Agentic engineering (Karpathy)

Foundations, level 2.

- **Definition:** Karpathy's February 2026 term: you orchestrate agents that write most code and act as oversight, and that is a learnable expertise.
- **Analogy:** From writer to editor-in-chief.
- **Say it:** Agents write most of the code; I design and verify the system they work in.

## Agent = model + harness

Foundations, level 1.

- **Definition:** The agent's behavior comes from the model plus everything wrapped around it.
- **Analogy:** The same engine in a race car and a tractor.
- **Say it:** When results disappoint, I check the harness before blaming the model.

## Guide (feedforward)

Context and harness, level 2.

- **Definition:** A harness control that steers the agent before it acts: AGENTS.md, skills, templates, reference docs.
- **Analogy:** Directions given before the trip.
- **Say it:** Guides prevent mistakes; sensors catch the ones that slip through.

## Sensor (feedback)

Context and harness, level 2.

- **Definition:** A check after the agent acts whose result goes back into its context so it self-corrects: tests, linters, review agents.
- **Analogy:** A smoke alarm the cook can hear.
- **Say it:** A sensor only counts if the agent sees its output.

## Computational control

Context and harness, level 2.

- **Definition:** A deterministic guide or sensor run by code: type checker, linter, test, blocking hook.
- **Analogy:** A ruler, not an opinion.
- **Say it:** Where code can judge, I use code.

## Inferential control

Context and harness, level 3.

- **Definition:** A guide or sensor where a model does the judging: rules files, reviewer agents, LLM judges.
- **Analogy:** A colleague's review.
- **Say it:** I use model judges only where code cannot decide.

## Builder harness

Context and harness, level 2.

- **Definition:** The part of the harness the agent's makers ship: system prompt, built-in tools, loop, compaction.
- **Analogy:** The car as it leaves the factory.
- **Say it:** I choose the builder harness; I engineer the user harness.

## User harness

Context and harness, level 2.

- **Definition:** The outer harness a team builds for its codebase: instructions, skills, tools, permissions, hooks, checks.
- **Analogy:** The fittings you add for your job.
- **Say it:** Most of my reliability gains come from the user harness.

## AGENTS.md

Context and harness, level 1.

- **Definition:** An open Markdown convention for project instructions read by many coding agents, now stewarded by the Agentic AI Foundation.
- **Analogy:** A README written for agents.
- **Say it:** AGENTS.md holds what every task needs, and nothing more.

## Agent Skills standard

Context and harness, level 2.

- **Definition:** An open format: a folder with SKILL.md plus optional scripts and references, loaded by progressive disclosure.
- **Analogy:** A manual on the shelf, opened only when needed.
- **Say it:** If only some tasks need it, it is a skill.

## Agentic AI Foundation

Loops and proof, level 3.

- **Definition:** A Linux Foundation body launched in December 2025 to steward MCP, AGENTS.md and goose.
- **Analogy:** Neutral ground for shared plumbing.
- **Say it:** MCP and AGENTS.md are vendor-neutral standards now.

## Structured notes

Context and harness, level 2.

- **Definition:** The agent writes progress and decisions to a file outside the window and rereads it later.
- **Analogy:** A lab notebook between shifts.
- **Say it:** Long work survives compaction because the plan lives in a file.

## Loop engineering

Loops and proof, level 1.

- **Definition:** Designing the goal, checks, stop conditions, state and budget so an agent keeps working toward a verified result without a human prompting each step. A young term.
- **Analogy:** Designing the thermostat, not turning the heater on and off.
- **Say it:** I stopped prompting the agent and started designing the loop that prompts it.

## Goal loop

Loops and proof, level 1.

- **Definition:** Repeating whole agent runs until a verifiable check passes or a limit stops it.
- **Analogy:** Rewrite until the editor approves, within a deadline.
- **Say it:** Fix until the suite is green, max ten rounds, max three dollars.

## Verified exit

Loops and proof, level 1.

- **Definition:** A stop condition checked by something other than the agent, such as a test run.
- **Analogy:** The referee blows the whistle, not the player.
- **Say it:** Done means a check passed, not that the agent said so.

## No-progress check

Loops and proof, level 2.

- **Definition:** Stopping a loop when its measure of progress stops improving for a set number of rounds.
- **Analogy:** Leaving a traffic jam that hasn't moved in ten minutes.
- **Say it:** Caps stop runaway loops; a no-progress check stops stuck ones early.

## Scheduled loop

Loops and proof, level 2.

- **Definition:** An agent run triggered on a timer or event, repeating across time.
- **Analogy:** A nightly patrol.
- **Say it:** Recurring work like triage runs as a scheduled loop with tight permissions.

## pass@k

Loops and proof, level 2.

- **Definition:** The chance that at least one of k trials succeeds. Measures capability.
- **Analogy:** Three tries to hit the target once.
- **Say it:** pass@k rises with k; it tells me what the agent can do.

## pass^k

Loops and proof, level 2.

- **Definition:** The chance that all k trials succeed. Measures consistency.
- **Analogy:** Hitting the target every single time.
- **Say it:** pass^k falls with k; it tells me whether users can rely on it.

## Grade the outcome

Loops and proof, level 3.

- **Definition:** Score what the agent produced, not the exact path of tool calls it took.
- **Analogy:** Judge the dish, not the order the chef chopped things.
- **Say it:** Path-checking evals punish valid creativity.

## Read the transcripts

Loops and proof, level 3.

- **Definition:** Reading actual runs to validate that graders are fair and passes are real.
- **Analogy:** Watching the game, not only the scoreboard.
- **Say it:** Before I trust an eval score, I read a sample of runs.