GLOSSARY
AI agent engineering terms
Working vocabulary used across the handbook. Where a term has a deeper page, the entry links to it.
By Oliver Wakefield-Smith, Digital Signet- Agent
- A system that can take actions toward a goal with some degree of decision-making latitude. The cornerstone definition lives on whatisanaiagent.com. [link]
- Agentic
- A property describing how much decision-making latitude a system has. Agentic is a continuum, not a category.
- Anti-pattern
- A pattern that solves the wrong problem or solves the right problem in a way that creates a worse one. Each pattern essay names the anti-patterns specific to that pattern. [read more]
- Autonomous
- A degree of agency where the system selects its tools and sub-agents without human approval per step. Highly agentic systems may still be human-supervised; autonomous systems are not.
- Confidence Gate
- Our coined name for a routing pre-check that compares model confidence against a threshold before the routing branch fires. Inaugural Pattern Deep Dive. [read more]
- Context window
- The number of tokens the model can attend to in a single inference. Long-running agents exhaust the window and hallucinate from the truncation point. Failure Pyramid level 5.
- Cost cliff
- Our coined name for the orchestrator-worker failure where the orchestrator decides to spawn N workers for a job that should have used three. Failure Pyramid level 2. [read more]
- Drift
- When an agent's output deviates from the intended outcome over time without raising an exception. The hardest failure mode to catch. Failure Pyramid level 1. [read more]
- Evaluator-optimiser
- Pattern where one model produces output and a second model critiques it and asks for a revision. Steady cost profile, prone to looping. [read more]
- Failure Pyramid
- Our coined taxonomy of the five categories of production agent failure ranked by frequency. [read more]
- Function calling
- A protocol for letting a model invoke a structured external function. The deterministic underpinning of tool use.
- LLM
- Large Language Model. The model class on which most 2026 agents are built. The cost line in any agent system is dominated by LLM tokens.
- Maturity Curve
- Our coined five-stage taxonomy of agentic deployment maturity. Self-assessment widget on the page. [read more]
- MCP
- Model Context Protocol. Anthropic's open protocol for connecting models to tools and data sources. The most-deployed tool-augmentation standard in 2026.
- Orchestrator-worker
- Pattern where one model decomposes a task into sub-tasks and dispatches them to worker models. The pattern we deploy most. Spike-prone cost profile. [read more]
- Parallelisation
- Pattern where the same task is run across multiple models or sub-agents in parallel for speed or vote-based reliability. [read more]
- Pattern
- A reusable structural shape for an agentic system. The five patterns from Anthropic's paper are the cornerstone vocabulary used here. [read more]
- Prompt chaining
- Pattern where the output of one prompt is the input to the next. Cheapest pattern at depth 3 or below; super-linear drift past that. [read more]
- Reflection
- A self-evaluation step where the agent critiques its own output before delivering it. A degenerate form of evaluator-optimiser.
- Routing
- Pattern where a classifier dispatches the input to one of several specialised sub-agents. The pattern most worth gating with a Confidence Gate. [read more]
- Sandbox
- An isolation boundary inside which an agent's tool calls execute. The presence or absence of a sandbox is the most material safety variable in any agent deployment.
- Silent drift
- Drift that completes without raising an exception. The most common failure mode in our pipeline. Failure Pyramid level 1. [read more]
- Tool
- Any callable function exposed to the model: HTTP request, file read, code execution, database query, search. Tools are the action surface of an agent.
- Tool router
- A sub-component, often a small fast model, that selects which tool to call next given a state and goal. Distinct from the orchestrator pattern.
- Tool cascade
- A failure mode where one tool returns an unexpected schema and downstream tool calls compound the error. Failure Pyramid level 4. [read more]
- Workflow
- A predefined path through model and tool calls. Distinct from an agent because the path does not change at runtime. Anthropic draws this line cleanly.
ABOUT THE AUTHOR
Oliver Wakefield-Smith
Founder, Digital Signet
Oliver runs Digital Signet, a research and product studio that operates ~500 production sites with AI agents as the engineering layer. The Digital Signet portfolio is built using a continuous AI-agent build pipeline, one of the largest agent-operated publishing operations on the open web. The handbook draws directly from those deployments: real cost data, real failure modes, real recovery patterns.