The AI agent glossary▌
Crisp, citable definitions of the terms that come up when you build with agents. Short enough to quote, precise enough to use.
Terms, defined plainly
- Agentic AI
- AI systems that plan, use tools, and take multi-step actions to complete tasks with limited human supervision.
- AI agent
- Software that perceives its environment, reasons, and acts toward a goal with a degree of autonomy.
- ASDLC
- The Agentic Software Development Lifecycle: BuildHaus's structured method for specifying, architecting, building, verifying, and shipping agent systems. read more
- Context window
- The amount of text a model can consider at once, measured in tokens.
- Embeddings
- Numeric vectors that represent the meaning of text, so similar content can be matched by similarity.
- Evals
- Evaluation tests that measure how reliably an agent completes a task before it ships.
- Fine-tuning
- Adapting a pre-trained model to a specific task by training it further on targeted data.
- Function calling
- An agent's ability to invoke external APIs and functions, so it can act rather than only generate text.
- Grounding
- Anchoring a model's output in retrieved, verifiable data rather than its training memory.
- Guardrails
- Controls that keep an agent's behaviour within safe, compliant, and predictable bounds.
- Hallucination
- A fluent, confident model output that is factually wrong or unsupported.
- Human-in-the-loop
- A design where a human approves or steers an agent's highest-stakes decisions.
- LLM
- Large language model: a model trained on text at scale to predict and generate language.
- Model-agnostic
- Built to work with any underlying model provider, with no lock-in to a single vendor.
- Multi-agent system
- Multiple agents that divide work, coordinate, and hand off to complete a task together.
- Observability
- The ability to trace and inspect what an agent did and why, so failures are diagnosable.
- Orchestration
- Coordinating the flow of work across tools, models, and agents to complete a task.
- Prompt
- The instruction and context given to a model that define the task it must perform.
- RAG
- Retrieval-augmented generation: grounding a model's answers in retrieved documents to reduce hallucination.
- Token
- The basic unit a language model reads: roughly a word or part of a word.
- Vibe coding
- Building software by prompting AI without engineering rigour, tests, or review.
- Workbench
- BuildHaus's open-source framework for agentic software development: project structure, agent workflows, and quality gates. read more
Asked often
What is an AI agent?
An AI agent is software that perceives its environment, reasons, and acts toward a goal with a degree of autonomy. It plans, calls tools and APIs, and carries out multi-step work toward an outcome, rather than only answering a single prompt.
How is an AI agent different from a chatbot?
A chatbot answers one message at a time. An agent plans a task, calls tools, and completes multi-step work toward an outcome. The short version: a chatbot responds, an agent acts.
How much does it cost to build an AI agent?
It depends on scope and risk. A validated prototype can cost a few thousand dollars; a production agent system with guardrails, evals, and handover typically runs from tens to hundreds of thousands. The better question is the cost of getting it wrong. Book a discovery call and we will scope it honestly.
How long does it take to build a production AI agent?
A scoped prototype takes weeks; a load-bearing production system takes months. BuildHaus works to the ASDLC so the timeline is predictable: specify, architect, build, verify, ship.
What is the ASDLC?
The Agentic Software Development Lifecycle: a structured method for shipping agent systems in five stages - specify, architect, build, verify, ship. It applies engineering discipline to agentic work so impressive demos do not become unmaintainable liabilities.
What is the difference between agentic AI and RPA?
RPA automates rigid, rule-based steps. Agentic AI handles variability: it reasons, adapts, and uses tools where the path is not fully scripted. RPA clicks buttons; agents make decisions.
$ book --type discovery --duration 30m --cost $0
Thirty minutes.
Zero dollars. Real answers.
Bring the problem, the half-built prototype, or just the hunch that agents could change how you work. We’ll tell you straight what’s worth building — and what isn’t.
Book a free 30-min discovery call-># no deck, no pressure, no obligation. exit code 0 either way.