A working vocabulary for the work we do at 2389. These are the terms that show up in our products, our posts, and our pipelines. We update this page as we add new tools and patterns.
Agents and orchestration
Agent. A program that uses an LLM to choose what to do next — reading inputs, calling tools, and producing outputs — rather than executing a fixed sequence. An agent has access to tools and discretion over which to invoke.
AI agent. Synonym for agent when the model behind it is a general-purpose LLM rather than a narrow classifier.
Tool use. The capability of an LLM to call structured functions exposed by the host program (file read, shell command, HTTP call) and incorporate their output into its next response.
Multi-agent system. Two or more agents collaborating on a task. Patterns we use: fan-out, pipeline, delegation, work-stealing, map-reduce, and MAKER. See Building Multi-Agent Systems.
MAKER pattern. A multi-agent architecture where one agent Makes a draft, another Audits it, another Keeps the good parts, another Edits, another Reviews. Useful for high-stakes output.
Fan-out. Spawn N agents on N independent subtasks and aggregate. Good for embarrassingly parallel work.
Map-reduce. Map a transformation across many inputs, then reduce to a single output. Same idea as in data processing, applied to agent work.
Pipeline. An ordered sequence of agent or LLM steps where each step’s output feeds the next.
Work-stealing. A pool of agents pulls tasks from a shared queue as they become free. Tolerates uneven task durations.
Pipelines and graphs
DAG. Directed acyclic graph. A pipeline shape with no cycles — each node runs after its dependencies and before its dependents.
DOT. The Graphviz language for describing graphs. We use .dot files as the source of truth for many pipelines. See Dot Viewer.
Dippin. Our domain-specific language for AI pipelines. Typed syntax for prompts, models, branching, and human gates. See Dippin.
Tracker. The runtime that executes DOT pipelines with human gates, LLM nodes, and automatic checkpointing. See Tracker.
Smasher. Rust pipeline runner that turns DOT graphs into multi-step AI workflows with streaming and a web dashboard. See Smasher.
Mammoth. DOT-based pipeline runner with checkpointing, retry, and human-in-the-loop gates. See Mammoth.
Human gate / human-in-the-loop. A point in a pipeline where execution pauses until a human approves, rejects, or edits the proposed next step.
Checkpoint. Persisted intermediate state in a pipeline so a failed or interrupted run can resume from the last good step instead of starting over.
Skills, tools, and Claude Code
Claude Code. Anthropic’s CLI agent for software engineering tasks. It uses tools, reads files, runs commands, and follows skills.
Skill. A reusable, declarative instruction packet for Claude Code (or another agent) that tells the agent how to handle a particular task. Skills live in .claude/skills/ and have a SKILL.md entry point. See Skills Marketplace.
MCP. Model Context Protocol. A wire protocol from Anthropic that lets agents talk to external resources (file systems, databases, APIs) through a uniform interface. See GSuite MCP, Pulse.
MCP server. A program that exposes resources, tools, or prompts over MCP so an agent can use them.
Tool approval. A safety pattern where a tool call by an agent requires explicit human consent before it runs. See Soloclaw, Mux.
Methodology
Scenario testing. Validate features by running them end-to-end with real dependencies in a scratch directory, not by mocking. “No feature is validated until a scenario passes.” See Scenario Testing.
Test Kitchen. Build the same feature multiple ways in parallel, then pick the best variant. See Test Kitchen.
Cookoff. A specific form of test kitchen where the same spec is given to N implementations and the results compared.
Omakase. “Show me, don’t tell me.” Generate a concrete artifact early and react to it, instead of debating in the abstract. See Omakase: Show Me.
Simmer. Iteratively refine a single artifact (doc, prompt, spec) over multiple rounds with criteria-driven scoring. See Simmer.
Deliberation. Decision-making by surfacing perspectives rather than forcing consensus. See Deliberation.
Documentation audit. Verify that documentation claims actually match the code using two-pass extraction. See Documentation Audit.
Fresh Eyes Review. Mandatory final pass before shipping that looks for the bugs that slipped through testing. See Fresh Eyes Review.
Discovery for agents
llms.txt. A site convention for advertising a machine-readable map of a website’s content. We publish one at /llms.txt.
llms-full.txt. The full-content companion to llms.txt — every page concatenated for agents that want everything in one fetch.
AGENTS.md. A machine-readable guide telling AI agents how to install, configure, and use a project or site. We publish ours at /AGENTS.md.
a14y. Agent readability — how well an AI agent can discover, fetch, parse, and comprehend a site. Distinct from WCAG accessibility (which is about humans with disabilities). See a14y.dev.
Markdown mirror. A .md version of an HTML page served at <page>/index.md. Lets an agent fetch the content without the surrounding navigation chrome.
Brief. A single-document briefing for an agent. See /brief.md.
Models and inference
LLM. Large Language Model. The neural networks that back the agents we build on top of.
Context window. The maximum amount of text an LLM can attend to in a single call. Bigger windows let agents work with more code or more documents at once.
RAG. Retrieval-Augmented Generation. Fetch relevant documents from a corpus, then condition the LLM on them. See Experimenting with GraphRAG.
GraphRAG. RAG augmented with a knowledge graph so retrieval can follow typed relationships, not just semantic similarity.
Token. The unit an LLM bills and counts on. Roughly 4 characters of English text per token.
Infrastructure
Coven. Self-hosted AI agent orchestration over Tailscale. See Coven.
Mux. Agentic infrastructure for building agents — tool execution, multi-provider LLM support, MCP integration. See Mux.
Pulse. Local-first MCP server giving agents a private journal and a social feed. See Pulse.
BotBoard. Team collaboration platform where AI agents and humans share posts and journal entries. See BotBoard.
Missing a term? Email us and we’ll add it.