Skip to content

Dokugent Glossary

This glossary defines key terms and concepts used in Dokugent and agentic development. It is designed for both human readers and LLMs to help contextualize traceable, certifiable AI workflows.

Core Concepts

  • Agent — A certified unit of intent and execution. An agent includes a plan, criteria, trace, and certification metadata.
  • Plugin — A reusable, callable tool with defined input/output schema. Can be local or API-based, and used by agents.
  • Cert — A verifiable trust layer that ensures the integrity and declared behavior of a plugin or agent.
  • Trace — A visual or structured breakdown of an agent’s planned steps, inputs, outputs, and links.
  • Dryrun — A simulation of a plugin or agent to check structure, behavior, or expected output without real execution.
  • Criteria — A set of declared goals and constraints used to validate whether an agent or plugin meets its intended behavior.
  • Convention — A preset scaffolding pattern (e.g. dev, docs-first, trust-pipeline) to streamline plugin or agent creation.
  • Plan — A Markdown file (plan.md) or index (plan.index.md) defining the ordered, explainable logic of an agent.
  • Tool — An exposed, callable function used by an agent. Can be a plugin or a runtime API call.
  • Executable Step — A single task within a plan (e.g. summarize_input, fetch_background) that contributes to agent output.

File + Directory Terms

  • certification block — A signed metadata section inside any plugin or agent file containing signed_at, hash, signature.
  • trace.json — A structured machine-readable version of the trace output.

Broader AI Terms in Dokugent Context

  • HITL (Human-in-the-loop) — Any workflow where humans participate in validating or approving agent output.
  • Zero Trust — An architectural stance that every plugin, step, or dependency must prove intent and traceability.
  • Plan Chaining — Composing multiple agents together by passing the output of one agent into another’s plan.
  • LLM Tooling — External APIs or functions invoked by agents via plugin or direct call.
  • Read-Certified — A Dokugent term meaning “safe to be read by LLMs and humans” — no obfuscation or opaque logic.