Glossary
Stable terminology for agentic engineering. Each entry includes a concise definition, plain language, technical depth, an example, and a common misconception.
- Agent
Software that repeatedly chooses and executes tool-mediated actions toward a goal under budgets and policies.
- Agent DNA
The declarative package describing an agent’s goal, tools/skills, policies, and evaluation hooks.
- Agent OS
An operating environment for building, running, and governing many agents with shared platform primitives.
- Agent runtime
The component that executes agent steps, mediates tools, enforces budgets, and persists run state.
- Checkpoint
Persisted run state at a safe point so execution can continue without restarting from scratch.
- Compensating action
A deliberate follow-up effect that reverses or mitigates a previously committed side effect.
- Context engineering
The practice of selecting, structuring, and budgeting what the model sees each step.
- Control plane
Systems that register, version, authorize, and operate agents and tools across an organization.
- Durable execution
An execution model where waits, retries, and recoveries are driven by persisted history or checkpoints.
- Evaluation (evals)
Offline and online methods to score whether agents meet task, safety, and cost goals.
- Guardrail
Preventive or detective control enforcing safety, policy, or quality constraints on agent behavior.
- Human-in-the-loop (HITL)
Design that inserts human judgment into an otherwise automated agent or workflow path.
- Idempotency
Property that performing the same logical operation multiple times yields one effect.
- Lease
A short-lived lock granting a worker exclusive rights to process a durable run.
- Memory
Persisted or retrieved information that informs future agent decisions outside a single model call.
- Model Context Protocol (MCP)
A protocol for discovering and invoking tools/resources from servers in a standard way.
- Reconciliation
Process of checking external side effects against intended step outcomes after ambiguity or failure.
- Skill
A reusable unit that groups instructions, tools, and constraints for a capability area.
- State
Structured data representing progress, decisions, and artifacts for a run or process.
- Tool
A schema-defined function or API adapter the runtime can execute on an agent’s behalf.
- Trace
A hierarchical record of operations for one agent run, typically model calls and tool calls.
- Workflow
An explicit sequence or graph of steps defined ahead of time, optionally including LLM nodes.