agentharnesses.dev

A field guide to agentic coding harnesses, for teams that have been told to adopt one.


Harnesses > Memory and context

Agent OS

Draft. This entry has not been checked against its sources yet.

Five slash commands that write your codebase's conventions into short standards files, load the relevant ones into a conversation or plan on request, and save each plan-mode spec to the repository.

MaintainerBrian Casel (Builder Methods)
CategoryMemory and context
HostsClaude Code
Components5 commands
LicenseMIT
Version3.0
Source codehttps://github.com/buildermethods/agent-os
Websitehttps://buildermethods.com/agent-os
Docsnone
DeepWikihttps://deepwiki.com/buildermethods/agent-os
Last researched2026-09-28, at commit 475b0ca

Features

Definitions are on the compare page.

FeatureSupportedNotes
Stages of work it covers
Brainstormingpartial/shape-spec asks what you are building and '1-2 clarifying questions if the scope is unclear', then about visuals, reference code, and standards. No step to compare approaches.
Written plansyes/shape-spec runs inside the host's plan mode, and task 1 of every plan saves it to agent-os/specs/ before any code.
Test firstnoNo test guidance. The test-first task loop from version 1 was retired with the implementation phases in 3.0.
Debugging processno
Code reviewnoNo review step. Version 2 had verifier subagents; 3.0 removed them.
Verificationno
Learning capturepartial/discover-standards writes conventions it finds in the code to agent-os/standards/, which later commands read. It records patterns on request, not lessons from finished work.
Branch to PRno
How it runs
Single entry pointnoFive separate commands with no dispatcher.
Automatic activationnoSlash commands only. Standards reach the agent only when you run /inject-standards or /shape-spec.
Session-start hookno
Blocking hooksnoNo hooks.
Subagentsno3.0 stopped installing subagents. /inject-standards can format standards for subagents you write yourself.
Parallel agentsno
Git worktreesno
Adopting it
Multiple hostspartialThe README says it 'Works alongside Claude Code, Cursor, Antigravity', but the installer writes only to .claude/commands/ and the commands call Claude Code's AskUserQuestion tool.
Per-project setupyesproject-install.sh copies commands and standards into the repository; commit them to share.
Team extensionspartialYour own standards and profiles, with inheritance in config.yml, are the extension point. No documented way to add or override commands.
No extra servicesyesBash scripts and Markdown. No network calls, keys, or accounts.

How it works

The workflow

Agent OS 3.0 keeps two jobs, standards and spec shaping, and hands the rest to the host. The changelog says it "refocuses the framework on what it does best—establishing and injecting standards", and that spec writing is "Now best handled using Plan mode", while the host's own todo lists and the model take over task breakdown and implementation. The implementation and orchestration phases of versions 1 and 2 are gone.

The five commands have no fixed order. The installer suggests starting with the first two below.

  1. Discover standards (/discover-standards). The agent proposes three to five areas of the codebase, reads "5-10 representative files" in the one you pick, and lists patterns that are unusual, opinionated, "tribal", or repeated. For each pattern you choose, it asks one or two questions about why the pattern exists, shows a draft, and writes the file only after you approve. Standards go to agent-os/standards/<folder>/<name>.md and are kept short: "Every word costs tokens."
  2. Inject standards (/inject-standards). With no arguments, it reads agent-os/standards/index.yml, matches the one-line descriptions against the conversation, and suggests two to five standards. With arguments (/inject-standards api/response-format) it loads those directly. It formats the result for one of three situations: implementation work (read the full text into the chat), writing a skill, or writing a plan (add @agent-os/standards/... references or paste the full text).
  3. Plan the product (/plan-product), optional. One question at a time about the problem, users, what makes it different, launch features, later features, and tech stack. Writes mission.md, roadmap.md, and tech-stack.md to agent-os/product/.
  4. Shape a spec (/shape-spec), which "must be run in plan mode". It asks what you are building, for visuals, for similar code to study, whether the work fits the product docs, and which indexed standards apply. It then builds a plan in which task 1 is always "Save spec documentation", and asks you to approve. The host carries out the rest of the plan.
  5. Index standards (/index-standards). Rebuilds index.yml after you add or delete standards by hand. It asks you to confirm a description for each new file and removes entries for deleted files without asking.

Components

How it steers the agent

Everything is wording in the commands. There are no hooks, reviewer gates, or scripts that check the agent's output.

Files it writes

Nothing is added to CLAUDE.md or AGENTS.md, so a new session knows nothing about the standards until a command loads them.

Install

Installation has two stages, and neither uses a plugin marketplace.

  1. Base install: a copy of the repository in your home folder, which holds your profiles. The documentation gives a clone command along the lines of git clone https://github.com/buildermethods/agent-os.git ~/agent-os; the docs site was not reachable from the research environment, so check the exact command at buildermethods.com/agent-os.
  2. Project install: from the project root, run ~/agent-os/scripts/project-install.sh. Use --profile <name> to pick a profile, or --commands-only to update the commands without touching standards. If agent-os/standards/ already exists, it asks before overwriting. Commit agent-os/ and .claude/commands/agent-os/ so the team gets the same setup.

The installer targets Claude Code only. The README says Agent OS "Works alongside Claude Code, Cursor, Antigravity, and other AI tools", but version 3 ships no install path or files for other hosts, and the commands name Claude Code's AskUserQuestion tool. Versions 1.4 to 2.x had a --cursor install flag.

At the researched commit, the default profile's example file sits at profiles/default/global/, while the installer copies only from profiles/<name>/standards/. A test install into an empty project copied the five commands and wrote an empty index.yml, but no standards. This does not matter if you build your standards with /discover-standards.

Needs bash and git. Version 3.0 had an install bug that made the script exit silently (discussion #311); the fix is in the unreleased changes included in the researched commit.

When to choose it

When not to

Version 3 replaced the commands of version 2 entirely; content from earlier versions (standards, specs, product docs) carries over, but tutorials for 1.x and 2.x describe commands that no longer exist.

Sources