Agent concepts
What an agent is, what's inside it, how it decides.
15 terms
Agent loop
The core runtime cycle of an agent: think, act, observe, repeat, until the goal is reached.
/glossary/agent-loop
Agent spawn
The runtime mechanism that lets one agent launch another, optionally in parallel, and wait for the result.
/glossary/agent-spawn
AI agent
An autonomous program that uses an LLM to plan and execute multi-step tasks via tools, in a loop, until a goal is reached.
/glossary/ai-agent
Compaction
A runtime process that summarises old turns once the context window is nearly full, freeing space for new turns.
/glossary/compaction
Coordinator agent
The top-level agent in a multi-agent setup. Plans, dispatches workers, merges their results.
/glossary/coordinator
Function calling
OpenAI's term for tool use. Same concept, slightly different format.
/glossary/function-calling
Goal pinning
A pattern where the original goal is auto-injected at the top of every turn so the LLM cannot lose the thread.
/glossary/goal-pinning
Multi-agent system
A set of agents that collaborate, typically a coordinator that spawns specialist workers in parallel.
/glossary/multi-agent
Persistent memory
State that survives across sessions, scoped per user or per agent.
/glossary/persistent-memory
Plan-first reasoning
A pattern where the agent emits a numbered plan before any tool call, then executes it step by step.
/glossary/plan-first
Read-before-edit
A safety guard that refuses to edit any file the agent has not first read, preventing hallucinated overwrites.
/glossary/read-before-edit
Specialist agent
An agent with a narrow, focused role spawned by a coordinator.
/glossary/specialist-agent
System prompt
The instruction text the LLM sees before any user message. Sets the agent's role, constraints, and behaviour.
/glossary/system-prompt
Tool use
The LLM capability to emit a structured call to an external function, which the runtime executes and feeds back as context.
/glossary/tool-use
Working memory
Per-session state the agent can write and read across turns, including a goal, a todo list, and arbitrary key-value notes.
/glossary/working-memory
Models & inference
LLMs, tokens, context, the math layer underneath.
9 terms
Context window
The maximum number of tokens an LLM can process in a single call. Modern frontier models offer 200K to 2M.
/glossary/context-window
Frontier model
The highest-quality, most expensive tier from a provider. Claude Sonnet, GPT-4o, Gemini Pro.
/glossary/frontier-model
Inference
The act of running a trained model to produce output. The thing you pay for per token.
/glossary/inference
LLM
A neural network trained on text that takes a prompt and returns text, optionally including structured tool calls.
/glossary/llm
Open-weight model
A model whose weights are publicly downloadable, runnable on your own hardware via Ollama, vLLM, or similar.
/glossary/open-weight-model
Streaming
Receiving the LLM's output token-by-token as it generates, instead of waiting for the full response.
/glossary/streaming
Temperature
A 0-to-1 dial that controls how random the model's output is. Lower = more deterministic.
/glossary/temperature
Tokens
The units LLMs process. Roughly four characters of English per token, billed per million.
/glossary/tokens
Tool format
The JSON shape a provider expects for tool definitions and tool calls. Differs slightly per provider.
/glossary/tool-format
RAG & knowledge
Retrieval, embeddings, the patterns for grounding answers.
8 terms
Chunking
Splitting source documents into smaller pieces (paragraphs, sections) before embedding them for retrieval.
/glossary/chunking
Embedding
A fixed-size numeric vector that represents the semantic meaning of a piece of text.
/glossary/embedding
Hybrid search
Combining vector similarity with keyword matching to get the best of both.
/glossary/hybrid-search
Knowledge base
The collection of documents an agent retrieves from. The 'data' side of RAG.
/glossary/knowledge-base
RAG
A pattern where relevant documents are fetched from a knowledge base and pasted into context before the LLM answers.
/glossary/rag
Re-ranking
Re-ordering retrieved candidates with a more expensive model to put the best matches at the top.
/glossary/rerank
Semantic search
Searching by meaning instead of keywords. Powered by embeddings and vector databases.
/glossary/semantic-search
Vector database
A database optimised for storing and querying high-dimensional vectors, typically for similarity search.
/glossary/vector-database
Runtime
Hooks, hot reload, triggers, the plumbing that makes agents shippable.
10 terms
Behaviour rules
Focused guards that protect against common agent failure modes (runaway loops, repeated tool calls, drift).
/glossary/behavior-rules
Cron trigger
A trigger that fires an agent on a cron schedule.
/glossary/cron-trigger
Filesystem module
The runtime module that exposes file operations (read, write, edit, grep, glob) to the agent.
/glossary/filesystem-module
Hooks
Declarative event handlers that fire at specific points in an agent's lifecycle (turn_start, tool_end, etc).
/glossary/hooks
Hot reload
Re-applying config changes to a running agent in place, without restarting the process.
/glossary/hot-reload
Module
A bundle of tools the agent can call (filesystem, web, shell, channels). Declared in YAML, loaded at boot.
/glossary/module
Triggers
Declarative entry points that launch an agent on an event: cron, webhook, file watch, or message.
/glossary/triggers
Webhook trigger
A trigger that exposes an HTTP endpoint, fires the agent when something POSTs to it.
/glossary/webhook-trigger
Workspace
The runtime module that mirrors the agent's writes to a live preview (React, LaTeX, slides, code, markdown).
/glossary/workspace
YAML-first config
Defining agents declaratively in a YAML file the runtime executes, instead of imperatively in framework code.
/glossary/yaml-first
Cost & ops
Routing, caps, compaction, audit logs.
7 terms
Audit log
A structured record of every agent turn, including tool calls, costs, and outputs, written to a sink.
/glossary/audit-log
Brain fallback
A second LLM provider the runtime swaps to when the primary returns an error or runs out of credit.
/glossary/brain-fallback
Cost cap
A hook that stops the agent when its session has exceeded a configured cost ceiling.
/glossary/cost-cap
Model routing
Sending different turn types to different models (cheap for exploration, premium for writing).
/glossary/model-routing
PII redaction
Pattern-matching personal data in tool results and replacing it before the LLM sees it.
/glossary/pii-redaction
Rate limiting
Capping a tool to N calls per minute, per session, or per user, to protect downstream APIs.
/glossary/rate-limiting
Turn cap
A runtime ceiling on how many loop iterations an agent can take before the runtime forces it to stop.
/glossary/turn-cap
Security
Credentials, scopes, encryption, the don't-leak-keys layer.
5 terms
Credential vault
An encrypted store for API keys, tokens, and connection strings, referenced from YAML by name.
/glossary/credential-vault
Envelope encryption
An encryption scheme where each row has its own data key, and that data key is wrapped by a master key.
/glossary/envelope-encryption
KMS
A managed service that holds master keys and performs cryptographic operations, never exposing the keys.
/glossary/kms
OAuth flow
A multi-step authentication flow where the user grants the agent scoped access to a third-party service.
/glossary/oauth-flow
Per-user scope
A credential resolution mode where each user brings their own key, resolved at session start.
/glossary/per-user-scope
Ecosystem
Frameworks, marketplaces, the wider field.
6 terms
Agent framework
Software that handles the agent loop, tool registration, and orchestration, so you do not write it yourself.
/glossary/agent-framework
Agent marketplace
A registry where agents can be published, discovered, and installed by name (like a package manager for YAML).
/glossary/agent-marketplace
Agentic
Adjective describing software that uses LLMs in a goal-directed loop with tools, instead of one-shot.
/glossary/agentic
CrewAI
A Python multi-agent framework that organises agents into role-based crews.
/glossary/crewai
LangChain
The most widely used Python framework for building LLM agents. Class-based, code-first.
/glossary/langchain
Self-hosting
Running your agent stack on infrastructure you control, with your own model provider keys.
/glossary/self-hosting