The Model

16 terms

AIA moving label, not a technology. Points at whatever computers can newly, impressively do — right now, large language models.->ModelThe parameters. Stateless — does next-token prediction and nothing else. Cannot do anything agentic on its own.->ParametersThe numbers inside a model — often billions — tuned during training. Everything the model knows lives in them. Also called weights.->TrainingThe process that sets a model's parameters by exposing it to vast amounts of text and adjusting to improve next-token prediction.->InferenceRunning a trained model to generate output — what happens on every model provider request. Parameters stay fixed.->EffortA dial for how much reasoning the model does before it answers. More effort spends more output tokens for a better shot at hard problems.->TokenThe atomic unit a model reads and writes. Roughly word-sized but not exactly. Context window size, cost, and latency all count tokens.->Next-token predictionWhat the model actually does. Samples one next token from the context, appends it, and runs again. Its only mode of operation.->Non-determinismThe same input can produce different output. A property of how models generate text and how providers serve requests.->Model providerWhatever serves a model for inference. Usually remote (Anthropic, OpenAI, Google), but can also be local (Ollama, llama.cpp).->Harness"Everything around the model that turns it into an agent: tools, system prompt, context-window management, permissions, hooks."->Model provider requestOne round-trip from the harness to the model provider. The harness sends context; the provider returns one response.->Input tokensTokens the harness sends on each model provider request. Billed at a lower rate than output tokens.->Output tokensTokens the model generates back. Billed at a higher rate than input tokens, since they cost more compute to produce.->Prefix cacheThe provider-side store that lets consecutive requests skip re-processing a shared prefix, billing those tokens at a lower rate.->Cache tokensInput tokens the provider has cached from a previous request via its prefix cache, billed at a much lower rate.->

Sessions, Context Windows & Turns

8 terms

StatelessCarries no information forward. The model is stateless across requests; an agent is stateless across sessions by default.->ContextThe relevant information the agent has access to right now — what the agent knows that's pertinent to the task.->Context windowEverything the model sees on each model provider request. Finite, model-specific, the only surface through which the model perceives.->StatefulCarries information forward. Sessions are stateful across turns; agents can be made stateful across sessions via a memory system.->AgentA model harnessed with tools, a system prompt, and a context window, that takes turns with a user. The model in motion.->System promptThe instructions the harness prepends to every model provider request — the agent's standing brief. Usually stable across a session.->SessionOne bounded run of interaction with an agent. Starts empty, accumulates, ends when cleared, closed, or compacted into a fresh session.->TurnOne user message plus everything the agent does in response, up until it yields back to the user. Contains one or more provider requests.->

Tools & Environment

10 terms

EnvironmentThe world the agent acts on — anything outside the harness that the agent perceives via tool results and changes via tool calls.->FilesystemA tree of files and directories the agent reads from, writes to, and executes within — the default environment for a coding agent.->ToolA function the harness exposes for the agent to call — Read, Write, Bash, Search. How an agent perceives and acts on the environment.->Tool callThe model's output naming a tool and its arguments — just structured text. The harness has to read it and execute.->Tool resultWhat the harness sends back after executing a tool call — file contents, output, or error. The agent's only view of the environment.->MCPA protocol for plugging external tool servers into a harness — how an agent gets tools beyond what the harness ships with.->Permission requestWhat the harness shows the user before executing a tool call that isn't pre-approved. The mechanism for putting a human in the loop.->Permission modeThe permission-gating slice of an agent mode — which tool calls trigger a permission request and which run automatically.->Agent modeA preset bundling a permission mode with behavioral instructions injected into the system prompt. Can flip mid-session.->SandboxAn isolated environment the agent runs inside — container, VM, or restricted shell. Limits the blast radius of agent actions.->

Failure Modes

9 terms

SycophancyConfidently agreeable model output. Caused by training that shaped the model to favor answers humans liked — including agreement.->Hallucination"Confidently-wrong model output. Two flavors: factuality (invented facts) and faithfulness (drift from loaded context)."->Parametric knowledgeWhat the model knows from training, stored in its parameters. Frozen at training time. Counterpart to contextual knowledge.->Knowledge cutoffThe date past which a model has no parametric knowledge. Post-cutoff libraries and APIs are fabrication traps unless docs are loaded.->Contextual knowledgeFacts the agent can read directly from the context right now. Counterpart to parametric knowledge.->Attention relationshipThe pairing between two tokens — meaningful pairs influence each other more than unrelated ones. A context of N tokens has ~N² of these.->Attention budgetEach token has a finite amount of influence to distribute across the rest of the context. Per-token, doesn't grow when context does.->Attention degradationAs a session grows, each token's attention budget spreads across more competitors; signal on meaningful relationships shrinks.->Smart zone"Early in a session the agent is sharp and focused. As the session grows it drifts into a dumb zone: sloppier, forgetful, more mistakes."->

Handoffs

9 terms

ClearingEnding the current session and starting a fresh one. The next message begins with an empty session and an empty context window.->HandoffTransferring agent context from one session to another, with no return path. Carry mechanism varies — artifact, compaction, others.->Primary sourceThe thing itself — code, transcripts, raw data. Complete and authoritative, but expensive to load into context.->Secondary sourceAn account of a primary source, one step removed — summaries, docs, compaction summaries. Cheap to load, lossy by construction.->Handoff artifactA document used as the carry mechanism for a handoff — written by one session to be read by another.->SpecA handoff artifact describing a multi-session piece of work — what's being built, not how each session does its share. Made of tickets.->TicketA handoff artifact scoping one session of work. Stands alone or hangs off a spec. Can block or be blocked by sibling tickets.->CompactionA handoff done in-memory: the previous session's history is summarised and seeds a fresh session. Lossy — detail traded for headroom.->AutocompactCompaction triggered automatically by the harness when the context window approaches full.->

Memory and Steering

6 terms

Memory systemA system that attempts to make an agent stateful across sessions by persisting to the environment and reloading at session start.->AGENTS.mdA file in the environment that the harness loads into the context window at session start — the project's standing brief to the agent.->Progressive disclosureLoading only the context an agent needs right now, with context pointers to the rest. Borrowed from UI design.->Context pointerA mention in one document that points to another, so the agent can pull it into context only when the task calls for it.->SkillA teachable capability bundled as a unit — kept out of the context window until a context pointer pulls it in for the task at hand.->SubagentAn agent spawned by another agent via a tool call. Runs in its own session, reports a single tool result. Cannot spawn further subagents.->

Patterns of Work

11 terms

Human-in-the-loopA working pattern where one or more humans pair with the agent during a session — reviewing, redirecting, or collaborating in real time.->AFKA working pattern where the user kicks off a session and leaves the agent to run unattended (away from keyboard).->Automated checkA deterministic verification that runs in the environment — tests, type checks, lints, build, pre-commit hooks. Pass/fail, no judgement.->Automated review"An agent reviewing another agent's work, often with a different model or system prompt. Non-deterministic: it forms a judgement."->Human reviewThe user reading the code the agent produced and forming a judgement on it. Reading the diff counts; reading the summary doesn't.->Vibe codingA working pattern where the user accepts the agent's code without human review. The diff is treated as opaque.->Design conceptThe shared understanding of what's being built, held in common between user and agent but separate from any asset.->Grilling"A technique for developing a design concept: the agent interviews the user Socratically, one decision at a time."->PrototypingHaving the agent build a quick, rough version when conversation is too low-fidelity and you need a real artifact to talk about.->DX"Developer experience: how easy a codebase and its toolchain make it for humans to do good work — docs, feedback speed, errors."->AX"Agent experience: how well the environment is set up for an agent to do good work — checks, architecture, and free context."->