How to Use This Glossary
The vocabulary of AI agents changes faster than the tools themselves. In 2024 "prompt engineering" was the headline term; by 2026 the conversation is about orchestration, subagents, context compression, and model routing — and a lot of writing uses these words loosely, or uses three words for the same thing.
This is a reference glossary of 30 terms, alphabetically grouped. Each entry is deliberately short, self-contained, and written to stand alone if an AI engine quotes a single sentence out of context. Where a term has a specific meaning inside Doforu, we say so in one clause rather than writing a second definition.
Read it front to back if you are new to the field, or jump to a letter range: A–C covers the foundations, D–H the pipeline mechanics, I–O the systems layer, P–S the workflow layer, and T–Z the coordinating concepts.
A–C
AI Agent
An AI agent is a software system in which a large language model (LLM) is given a goal, a set of tools, and the ability to take multi-step actions toward that goal — reading files, running commands, calling APIs — rather than only producing a text response. The defining property is autonomy inside a loop: the agent observes the result of an action and decides what to do next. A chatbot answers; an agent acts.
Agent Orchestrator
An Agent Orchestrator is the coordinating layer that decomposes a goal into tasks, assigns those tasks to agents, sequences what must run in order, runs in parallel what does not, and merges the results into one deliverable. It is the difference between "an agent" and "a team of agents with a manager." Doforu is built as a desktop Agent Orchestrator: it splits a goal and dispatches multiple subagents simultaneously, whereas most competing tools run a single agent sequentially.
Agentic AI
Agentic AI is the broad label for AI systems that plan, use tools, and execute multi-step work with limited human intervention, as opposed to systems that answer a single prompt and stop. It is an umbrella term covering agents, orchestrators, and multi-agent systems — useful for describing a category, too vague to specify an architecture.
Air-Gapped
Air-gapped describes a machine or network that is physically isolated from the internet and from any other network. For AI tooling it is the strictest form of local operation: the model weights, the agent runtime, and your code all live on isolated hardware, and no request can leave the environment. Doforu can be used air-gapped by pairing it with local models via Ollama or vLLM, since orchestration runs entirely on the machine.
Autonomous Agent
An autonomous agent is an agent that can plan and execute a multi-step task without human approval at each step, deciding its own next action until it either completes the goal or hits a stopping condition. Autonomy is a spectrum, not a binary: most production systems are "autonomous with guardrails" and a human checkpoint at defined boundaries.
BYOK (Bring Your Own Key)
BYOK means the user supplies their own API key for a model provider, and the provider bills the user directly at its published rates, with the tool adding no token markup. It separates the software vendor's revenue from your inference spend, which means your cost tracks actual usage and you can switch providers without changing tools. Doforu is BYOK by design and also supports fully local models, where there is no per-request cost at all.
Context Compression
Context compression is the set of techniques — summarizing, truncating, or selectively dropping earlier conversation and tool output — that keep a long-running agent task inside the model's context window without losing what matters. It is what allows an agent to work for hours or across a large codebase rather than losing the thread every few turns. Doforu applies context compression automatically as agent sessions grow, so long orchestrations do not degrade or fail on window limits.
Context Engineering
Context engineering is the discipline of deciding what information an LLM should see at each step — retrieved documents, file contents, tool schemas, prior results, system instructions — and in what form and order. It is the successor framing to prompt engineering: the hard problem in 2026 is usually not wording a request well, but assembling the right state for the model. The term was popularized through 2025 writing on agent reliability and by mid-2026 is standard vocabulary in agent documentation.
Context Window
The context window is the maximum amount of text, measured in tokens, that a model can consider at once — the working memory for a single inference. Everything the model "knows" about your specific task must fit inside it: system prompt, retrieved context, conversation history, tool output, and its own reply being generated. When agent tasks exceed the window, you need retrieval, compression, or task decomposition.
D–H
Embedding
An embedding is a numerical vector that represents the meaning of a piece of text so that similar meanings sit close together in vector space. Embeddings power semantic search, retrieval, clustering, and de-duplication, and they are the substrate of RAG. They are usually produced by a separate, smaller model than the one doing the reasoning.
Fine-Tuning
Fine-tuning is further training a pretrained model on your own examples so its weights adapt to a specific task, tone, or domain. It changes the model itself, unlike RAG or prompting, which change only what the model sees at inference time. Fine-tuning is powerful but expensive to maintain, and for most agent use cases in 2026, good context engineering delivers more value per unit of effort.
Function Calling (Tool Calling)
Function calling, also called tool calling, is the mechanism by which a model emits a structured request — a function name and JSON arguments — that the surrounding software executes, with the result fed back to the model. It is how agents touch the world: reading files, running shell commands, querying databases, or hitting an API. Nearly all serious agent frameworks in 2026 are built on this primitive.
Guardrails
Guardrails are the constraints that bound what an agent is allowed to do: permission lists, allowlisted commands, directory restrictions, approval requirements, spending caps, and output filters. They are what make autonomy safe enough to deploy, and they sit outside the model — a guardrail is enforced by the system, not requested in a prompt. An agent without guardrails is a demo, not a product.
Hallucination
A hallucination is a confident, fluent output that is factually wrong or unsupported by any source — an invented API, a nonexistent file path, a fabricated citation. It is not lying and not a bug in the ordinary sense; it is a consequence of models generating plausible text rather than verifying claims. Grounding (retrieval), tool use (actually checking the file), and verification steps are the practical mitigations.
Human-in-the-Loop
Human-in-the-loop (HITL) describes a workflow where a person reviews or approves specific agent actions at defined checkpoints rather than supervising every step or none at all. Good HITL design places checkpoints where errors are costly and irreversible — applying a migration, pushing to main, sending an external message — and stays out of the way everywhere else.
I–O
Inference
Inference is the act of running a trained model to produce an output from an input — one forward pass, or a sequence of them, through the network's weights. It is distinct from training: no weights change. Inference is what you pay for per token with a hosted API and what consumes GPU or CPU time when you run a model locally.
LLM
A large language model (LLM) is a neural network trained on very large text corpora to predict the next token, which at scale yields general-purpose language understanding, reasoning, and code generation. It is the engine inside an agent — the part that decides — but an LLM alone is not an agent: the agent adds tools, memory, a loop, and stopping conditions.
Local-First
Local-first describes software where the core computation and the primary copy of your data live on your own device, with the cloud as an optional extra rather than a requirement. Applied to AI coding, it means the agent runtime and your code stay on your machine; if a cloud model is used, only the specific prompts you send leave. Doforu is local-first: orchestration, file access, and execution history stay on your desktop.
MCP (Model Context Protocol)
MCP (Model Context Protocol) is an open standard that defines how AI applications connect to external tools and data sources through a uniform interface, so a tool built once works across many agent clients instead of being rewritten per vendor. It is often described as a USB-C port for AI integrations, and it has become the common integration layer across agent tooling in 2026. Doforu uses MCP as its integration mechanism, so external tools and services plug in without custom glue per tool.
Model Routing
Model routing is the practice of sending different tasks to different models according to cost, speed, capability, or privacy requirements — a cheap fast model for classification and file summarization, a frontier model for architectural reasoning, a local model for anything that must not leave the machine. Routing is where a lot of practical cost control in agent systems actually happens. Doforu supports routing across 15+ model families, including local models via Ollama and vLLM.
Multi-Agent System
A multi-agent system is an architecture in which several distinct agents, each with its own role, tools, and context, cooperate on a task rather than one agent doing everything. The motivation is context isolation and specialization: a reviewer agent that sees only the diff is more useful than one drowning in the full history. The cost is coordination overhead, which is why orchestrators exist.
Orchestration
Orchestration is the act of coordinating multiple components, agents, or steps so that they execute in a coherent order and produce a combined result. In agent systems it covers task decomposition, scheduling, dependency management, retries, and result aggregation. It is a control-plane concern: orchestration decides who works, in what order, and with what inputs — it does not call the model for you.
P–S
Parallel vs Sequential Execution
Sequential execution runs agent steps one after another, each waiting for the previous to finish; parallel execution runs independent steps at the same time. Sequential is simpler and easier to debug; parallel is far faster on tasks that decompose naturally, such as touching twenty files, but requires dependency tracking and result merging. Doforu's Orchestrator 2.0 dispatches multiple subagents simultaneously, which is the structural difference from tools that run one agent turn at a time.
Plan / Agent / Orchestrator Modes
Plan, Agent, and Orchestrator are Doforu's three operating modes, matched to how much autonomy and coordination a task needs. Plan mode produces a proposed approach for you to review before anything is changed; Agent mode executes a task directly with a single agent; Orchestrator mode decomposes a goal and runs multiple subagents in parallel with dependency-aware scheduling.
Prompt Engineering
Prompt engineering is the practice of writing instructions to a model — role framing, examples, constraints, output format — so it reliably produces what you want. It remains useful and is now one part of the broader discipline of context engineering: in an agent, the prompt is a small fraction of the tokens the model actually sees.
RAG (Retrieval-Augmented Generation)
RAG (Retrieval-Augmented Generation) is a pattern in which relevant documents are retrieved from an external store and inserted into the prompt, so the model answers from supplied evidence instead of its trained parameters. It is the standard fix for both hallucination and stale knowledge, because updating the retrieval index updates the system's knowledge without retraining anything. It is central to code assistants, where "relevant documents" means the right files from a large repository.
Reranking
Reranking is a second-stage scoring pass that reorders a set of candidate results retrieved by a fast but imprecise search, typically using a more expensive cross-encoder that reads the query and each candidate together. Because only the top few results reach the model's context, reranking improves answer quality at low token cost and is one of the highest-leverage refinements in a RAG pipeline.
Sandbox
A sandbox is an isolated execution environment where an agent's code can run — a container, virtual machine, or restricted process — so that mistakes, malicious output, or destructive commands cannot reach your host system or network. It is the containment half of guardrails: guardrails decide what is permitted, the sandbox limits the blast radius when something gets through. Doforu runs agent actions locally with explicit permissions, and sandboxes risky execution.
Skill (SKILL.md)
A Skill is a reusable, declarative capability definition — in Doforu's case a SKILL.md file — that tells an agent how to perform a specific, repeatable job, written in natural language and shared like any other project file. The point is amortization: describe a workflow once, and every agent and every teammate reuses it instead of re-explaining. Doforu's Skills are team-shareable, so a working definition of process becomes an asset rather than a personal prompt.
T–Z
Subagent
A subagent is a secondary agent spawned by an orchestrator to handle one bounded piece of a larger goal, with its own task, its own context, and usually its own result to report back. Because each subagent sees only what its task requires, work stays coherent instead of collapsing into one overloaded context. Doforu's subagents run in parallel — several work at once and the orchestrator merges their output.
Learn More
If this glossary answered "what does that word mean," the next question is usually "what does the architecture look like in practice." These pages go deeper:
- What is an Agent Orchestrator? — the full explanation of the coordinating layer, and how it differs from a single coding assistant.
- Getting Started with Multi-Agent Development — a walkthrough of defining, running, and monitoring a parallel agent workflow.
- MCP Integration — how Doforu connects to external tools and data sources through the Model Context Protocol.
- Skills Guide — how to write a SKILL.md so a workflow is reusable and team-shareable.
- Doforu vs Cursor — an honest comparison of a parallel orchestrator against a leading AI-first editor.
Final Thoughts
Terminology is only worth learning when it changes decisions. Two of these terms — orchestration and context engineering — account for most of the difference between agents that are impressive in a demo and agents that are useful on Monday morning. The rest are vocabulary you need to evaluate tools without being sold an architecture you cannot inspect.
One last note on definitional honesty: this field has no standards body, and vendors routinely redefine terms to favor their own design. Where a term has a contested meaning, the definition above follows the most common technical usage rather than any single product's marketing. If a vendor's definition of "agent," "orchestrator," or "autonomy" conveniently describes only their product, that is worth noticing.
Ready to try a local-first desktop agent orchestrator? Download Doforu for macOS, Windows & Linux
