Reference
Every term here reduces to something simpler. When a word starts to feel vague, look up what it reduces to.
= a standard way for one agent to hand a task to another agent over HTTP, without seeing its insides
Each agent publishes a card that says what it can do. A piece of work is a task with an id and a status. MCP connects an agent to tools, and A2A connects agents to each other. Chapter 13
= a runtime loop + a model that chooses the next action + tools + stop rules
In a workflow, code decides what happens next. In an agent, the model does, within limits the runtime enforces. Chapter 0
= while loop: build context, call model, run requested tools, append results
The loop ends on a final answer, a budget, an error rule, or a person. Chapter 1
= a README for coding agents: build steps, test commands, and conventions, read into the context
A repository can have several. When their instructions conflict, the file closest to the code being edited wins, and the rest still apply. Chapter 9
= intercept a matching call + checkpoint + ask a person + resume with their decision
The approval is bound to one exact action. Edited or replacement actions are checked again. Chapter 11
= a shared workspace where specialists read and post partial results
Coordination happens through the shared store rather than direct messages. Chapter 13
= a durable snapshot of run state after a step
It holds enough to resume: history, plan, step, pending actions. Chapter 12
= the model writes a short program instead of a single tool call + a sandbox runs it
One action can call several functions, loop, and compute. It needs strong containment. Chapter 2
= replace old context with a shorter form (summary, placeholders, references)
It keeps a long run inside the window. Originals can stay in storage. A careless summary can drop constraints. Chapter 4
= limits on what the process can touch at all: files, network, devices, resources
Enforced by the OS, a container, a VM, or the network. Its strength depends on configuration. Chapter 10
= the exact input to one model call, rebuilt every call
Instructions, tool definitions, conversation, tool results, retrieved material. A view of state under a token budget. Chapter 4
= misleading or distracting content in the context
Irrelevant files, stale facts, failed attempts, or injected instructions. Chapter 4
= the maximum tokens one call can take, input plus output
A hard limit. Quality can drop well before you reach it. Chapter 4
= keep long-lived secrets out of the agent's reach; hand out scoped, short-lived tokens
If there is nothing valuable to steal, exfiltration does little harm. Chapter 10
= a tool call that starts another agent loop with a task spec
The spec says the objective, scope, tools, output format, budget, and what the child owns. Chapter 7
= an agent with recovery guarantees: resume after crashes without losing work or repeating side effects
Defined by what it can recover from, not by how long it runs. Chapter 12
= a list of network destinations the process may reach
It blocks exfiltration to unknown hosts. An allowed host can still carry data out. Chapter 10
= stored records of past tasks + retrieval by similarity to the current task
What happened, what worked, what failed. Chapter 5
= a handler that loads state, runs the loop, and saves state, once per event
Messages, webhooks, schedules, or file changes start runs. Chapter 1
= a rule enforced by the other system, whatever the agent does
Branch protection, scoped tokens, access policies, deduplication on the receiver. Chapter 10
= a state machine whose nodes may call a model
You define the process shape. The model works inside nodes or picks some edges. Chapter 14
= transfer control of the conversation to another agent
Unlike agents-as-tools, the caller does not get control back with a result. Chapter 7
= the ordinary code around the model: loop, context builder, policy, tool dispatch, state
Control and authority live here. Chapter 0
= delegation more than one level deep
Directors, managers, workers. Each level adds coordination cost and summary loss. Chapter 14
= doing it twice has the same effect as doing it once
For outside effects, the receiver must deduplicate on a key that you saved before the first attempt. Chapter 12
= text asking the model to behave a certain way
It depends on the model following it. Injected text can argue against it. Chapter 10
= a model call that grades an output against a rubric
Cheap and flexible. Calibrate it against human grades before you trust it. Chapter 16
= one agent splits work, subagents do the parts, the manager combines results
Also called orchestrator and workers. Chapter 14
= RPC for listing and calling tools, resources, and prompts in another process
It standardizes the plumbing between an agent and tool servers. It is not an agent loop and not a security boundary. Chapter 2
= storage + a write policy + retrieval + injection into context
Not the model remembering. Chapter 5
= a function from one context to generated tokens
It runs no code and keeps nothing between calls. Chapter 0
= a message with the outcome of a tool call, tied to its call id
Output, error, or "rejected". Chapter 1
= intended, attempted, confirmed, outcome unknown
Track them for every outside side effect so recovery knows what to do. Chapter 12
= generate and maintain an explicit list of intended future actions
Text, a todo structure in state, or a separate planner's output. Chapter 6
= one call writes a plan + a loop carries out steps + replanning on failure
= short descriptions always in context + full content loaded only when needed
The usual way skills are loaded. Chapter 8
= the provider reuses computation for an identical prefix
Lowers cost and latency for the cached part. Does not free context space. Reuse depends on provider rules. Chapter 4
= instructions hidden in data the model reads
A web page, file, or email that steers the agent. Defend outside the model. Chapter 10
= the API in front of the model: may store conversations and cache prefixes
Each call still conditions only on that call's context. Chapter 0
= alternate model reasoning and tool actions, one step at a time
The simple loop most agents start from. Chapter 1
= another model call that critiques earlier output or state
Best with an outside signal such as tests. It costs calls and can repeat the same mistake. Chapter 6
= query + filters + candidates + selection under a budget + provenance check + injection
Similarity is not relevance. Relevance is not correctness. Chapter 5
= one task execution, one model call plus its tools, one user exchange, a saved conversation, the recorded history
= code in the tool path that blocks or pauses matching actions
Enforced on the paths it intercepts. An unintercepted path goes around it. Chapter 10
= containment applied to the agent's tools
A container, VM, or OS sandbox with limited files, network, and credentials. Chapter 10
= stored items retrieved by relevance to the current need
Often vector search, often combined with keywords. Chapter 5
= a reusable package selected at runtime: instructions, tools, a workflow, a script, knowledge, or a learned procedure
What it reduces to depends on the system. Chapter 8
= what the system tracks to do its work: conversation, runtime variables, the world, saved data
Context is a view of it. Chapter 3
= a small curated file or table of facts, usually injected into every context
"Project uses PostgreSQL." Chapter 5
= model output that must match a schema: constrained decoding, or validate and retry
= another agent loop with its own context, started by a parent, returning a result
= a list of tasks with statuses in runtime state, updated by a tool, often shown to the model each turn
It keeps the plan where later steps can read it. Saved to disk, it also survives a restart. Chapter 6
= name + description + input schema + implementation + policy metadata
= a structured request from the model: tool name + arguments + call id
A request, not an execution. Your code decides whether to run it. Chapter 2
= private data + untrusted content + a way to send data out
An agent with all three can be steered into leaking data. Remove one leg. Chapter 10
= checking evidence of success: tests, API results, a person, or a separate checker
Different from the model proposing it is done. Chapter 0
= a code-chosen sequence of steps, some of which call a model
Often the right choice when the steps are known. Chapter 0
= whatever is in the current context