AI
#The label keeps moving. It points at whatever computers can newly do that impresses us, and right now that means large language models.
AI Coding Dictionary
The vocabulary of AI coding, in plain English. Skimmable definitions for the terms that make agentic coding click.
/
No terms found
Nothing matches this search. Clear the query or pick a different section.
01 The model · 6 terms
What the model is, and the units it thinks in.
Model
#Frozen numbers that predict the next token. A model on its own holds no state. It becomes an agent only once a harness hands it tools and memory.
Parameters
#The billions of tuned numbers inside a model. Everything it knows from training lives in them. Also called weights.
Training
#The process that sets a model's parameters by showing it vast amounts of text and nudging it toward better next-token prediction.
Token
#The atomic unit a model reads and writes, roughly word-sized but not exactly. Context size, cost, and latency are all counted in tokens.
Model provider
#Whatever serves a model for inference. Usually that means a remote API from Anthropic, OpenAI, or Google, and sometimes a local runtime like Ollama or llama.cpp.
02 Sessions, context and turns · 8 terms
How an agent sees the world, one request at a time.
Agent
#A model in motion, working toward a goal across turns. The harness, the tools, the system prompt, and the context window are all part of it.
Context window
#Everything the model sees on one provider request. It is finite, it differs by model, and it is the only way the model perceives anything at all.
Context
#The relevant information the agent has right now. It is the slice of what the model knows that matters for the task in front of it.
System prompt
#The standing instructions the harness puts in front of every request. It works like the agent's brief, and it usually stays the same for a whole session.
Turn
#One user message plus everything the agent does in response, until it yields back. A turn can contain several provider requests.
Session
#A run of turns that share one context window. It ends by clearing, and the next message starts from an empty window.
Context engineering
#The discipline of deciding what deserves a slot in the finite context window, for how long, and in what form. What the model gets to see matters more than what you say to it.
Subagent
#An agent spawned by another agent with a fresh context window and a narrow brief. It digs, then returns only the conclusion, so the parent window never carries the search.
03 Tools and environment · 10 terms
How an agent perceives and changes the world.
Tool
#A function the harness exposes for the agent to call, like Read, Write, Bash, or Search. Tools are how an agent touches its environment.
Tool call
#The model's output naming a tool and its arguments, just structured text. The harness has to read it and actually execute it.
Harness
#Everything around the model that turns it into an agent: tools, system prompt, context management, permissions, hooks.
MCP
#The open protocol for plugging external tool servers into a harness, which is how an agent gets tools beyond the ones the harness ships with.
Permission request
#What the harness shows before running a tool call that isn't pre-approved. The mechanism that keeps a human in the loop.
Sandbox
#An isolated environment the agent runs inside: container, VM, or restricted shell. Limits the blast radius of whatever it does.
Skill
#Reusable know-how the agent loads on demand, usually a Markdown file that teaches a process and stays separate from the data it acts on.
Progressive disclosure
#Load the map first, the territory on demand. The agent sees pointers and fetches full content only when relevant, so the window carries summaries instead of libraries.
Hook
#Code the harness runs at a fixed point in the loop, no matter what the model decided. Hooks turn requests like 'be careful' into enforcement that cannot be talked away.
Guardrail
#Any constraint that keeps agent behavior inside bounds the human chose. Permissions, sandboxes, and hooks are all guardrails; prompting is not.
04 Failure modes · 9 terms
The ways models go wrong, and why sessions decay.
Hallucination
#Confidently wrong output. It comes in two flavors. Factuality means inventing facts that were never there, and faithfulness means drifting away from the context the model was given.
Sycophancy
#Confidently agreeable output. The model tells you your idea is great when it isn't, because training rewarded the answers humans liked.
Knowledge cutoff
#The date past which a model has no parametric knowledge. Post-cutoff libraries and APIs are fabrication traps unless you load the docs.
Parametric knowledge
#What the model knows from training, frozen in its parameters. The counterpart to contextual knowledge you load in at runtime.
Attention budget
#Each token has a finite amount of influence to spread across the rest of the context, and it doesn't grow when the window does. Long sessions dilute focus by construction.
Smart zone
#Early in a session the agent is sharp. As context grows it drifts into a dumb zone where it gets sloppier and starts forgetting things. Clear the session before you get there.
Prompt injection
#Malicious instructions smuggled in through content the agent reads: a web page, an issue comment, a tool description. The agent treats data as orders because nothing in the window marks which is which.
Confused deputy
#An agent tricked into misusing its own authority. It follows hostile instructions it found in data while believing it serves you, because data and orders arrive through the same channel.
Tool poisoning
#Prompt injection aimed at tool definitions. A compromised tool description steers every agent that loads it, so one poisoned server turns all of its users into delivery mechanisms.
05 Handoffs and memory · 7 terms
Moving context between sessions without losing the plot.
Handoff
#Moving agent context from one session to another, with no return path. The carry mechanism varies: artifact, compaction, or a fresh brief.
Clearing
#Ending the current session and starting fresh. The next message begins with an empty session and an empty context window.
Compaction
#A handoff that happens in memory. The old session gets summarized to seed a fresh one, trading detail for headroom.
Spec
#A handoff artifact describing a multi-session piece of work. It says what is being built, not how each session does its own share, and it is made of tickets.
Ticket
#A handoff artifact scoping one session of work. Stands alone or hangs off a spec, and can block or be blocked by sibling tickets.
Memory system
#Any system that makes an agent stateful across sessions by persisting to the environment and reloading at session start. AGENTS.md is the simplest one.
Checkpoint
#Saved state a run can rewind to after a crash or a bad turn. Usually commits plus conversation records, so recovery never means redoing hours of work.
06 Patterns of work · 7 terms
The working patterns that make agentic coding click.
Human-in-the-loop
#A working pattern where a human pairs with the agent during the session, reviewing and redirecting in real time.
Vibe coding
#A working pattern where you accept the agent's code without review, so the diff goes unread. It is fast, and it leaves the agent's mistakes in place.
Grilling
#A technique for sharpening a plan. The agent interviews you Socratically, one decision at a time, before anything gets built.
Prototyping
#Having the agent build a quick rough version when conversation is too low-fidelity and you need a real artifact to argue with.
Automated check
#A deterministic verification that runs in the environment: tests, type checks, lint, build. Pass or fail, no judgment involved.
Human review
#You reading the code the agent produced and forming a judgment. Reading the diff counts, and reading the summary doesn't.
Steering
#Talking to an agent while it works, as course corrections at safe points rather than joystick control. Keep it for the moments when two minutes of input saves two hours of confident wandering.
07 Retrieval and knowledge · 8 terms
How agents find what they know: vectors, search, and reranking.
Embeddings
#Numbers that place meaning in space. Texts with similar meanings land near each other as vectors, and every semantic search ranks that geometry.
Vector database
#A database built to store embeddings and answer one question fast, namely which stored items sit closest to this query vector. The index runs the search and metadata filters narrow the set.
ANN
#Approximate Nearest Neighbor, the trade behind fast vector search. Skip the exact ranking, return something very close, and spend the time you saved elsewhere.
HNSW
#The approximate index most vector databases reach for. It builds a layered graph where each hop lands closer to the query, trading a sliver of accuracy for orders of magnitude in speed.
Hybrid search
#Keyword search for precision plus vector search for paraphrase, merged into one ranking. Each covers what the other misses.
Reranking
#A second, stronger pass over the top candidates. Score each query and document pair carefully, then reorder. Recall broadly first, then spend compute where it changes the order.
Reciprocal rank fusion
#A way to merge ranked lists without comparing scores. Each document earns points from where it sits in every list, which is why hybrid search rarely needs a learned mixer.
Cosine similarity
#A distance metric that compares the angle between vectors and ignores their length. Document length stops mattering, direction is everything.