Product
: What a person uses on top of the stack; four products since ADR 0106.
Operator: The person who gives the stack an intent or a story and runs teams of agents on real work.
Agency (proposed): The product control-plane converges into, where the operator manages agents, work, teams, auth and tools, each definition holding a loom.agent/1 spec by value.
loom-coder (proposed): A Loom-based coding agent with a command line and a web version, whose repository is not yet known.
swarm (paused): A goal loop that drives Claude Code through metaharness directly, without Loom, and is frozen.
agentide: A local coding workbench where humans and agents share one durable session.
control-plane (paused): A local workspace service for goals, plans, implementation assignments and verified merge receipts, now paused.
Application
: An end-to-end use of the stack; ADR 0096 has applications build on loom-sdk.
conductor: Coordinates Claude Code sessions across an organization's repositories, starting one controller session per repository.
takt: Delivers stories on loom-sdk and counts nothing done until a validator records evidence bound to a commit.
b10x-loom: Loom's command line and plugins, an application built on loom-sdk.
Agent runtime
: Opens a case on a protocol, governs it and invokes effects; calls llm for decisions and metaharness for agent work.
Loom: The agent runtime and owner of the loom.agent/1 spec, which opens a case on a protocol, admits only the case's actions and serves A2A where a spec exposes it.
loom-sdk: Loom's library, on which applications run their cases with Commission's runtime and Loom's governor.
Commission: Loom's case runtime: case, frontier and authority, which rechecks each proposed action and invokes its effect.
a2a: The A2A protocol contract as an ESS specification and the crate generated from it, a library Loom's host serves agents through and not a layer.
Meta-harness
: Starts any harness behind one event stream and one command set, and records every session as trace-ir/1.
metaharness: Starts any harness behind one event stream and one command set, and allows, denies or rewrites each tool call.
Harness
: The agent loop and its tools: turns, prompt assembly, tool round trips, compaction, budgets and approvals.
harness: The in-house agent loop: turns, prompt assembly, tool round trips, compaction, session and transcript, budgets and approvals.
Claude Code: A vendor harness that metaharness drives and that calls Anthropic's Messages API.
Codex: A vendor harness that metaharness drives and that calls OpenAI's Responses API.
Pi: A vendor harness that metaharness drives through the same protocol as Claude Code and Codex.
OpenCode: A vendor harness that metaharness drives through the same protocol as Claude Code and Codex.
tools: Tool contracts and their default implementations, which harness takes its tools from.
Model call
: One turn in, one turn out: provider wires, credential presentation (custody and OAuth2 refresh in Connectors, ADR 0107), routes, usage and cost.
llm: One turn in, one turn out: provider wires, credential presentation, routes, usage and cost, and not an agent loop.
inference: Composes speech-to-text, language models and text-to-speech into streaming graphs, its language model node wrapping llm.
llm-gateway: The serving side of llm, through which self-hosted models are served.
Model providers
: Hosted and self-hosted model endpoints, outside beyond10x.
Anthropic: A hosted model provider reached over the Messages API.
OpenAI: A hosted model provider reached over the Responses API.
self-hosted: Self-hosted model endpoints such as vLLM, served through llm-gateway.