Get started

Install a supervised agent workflow matched to the work you need to do.

Start with a real task whose result you can verify. Most software teams should start with core: it is the catalogue’s flagship build system, not merely a sample. Product, design, infrastructure, research, and catalogue work remain first-class starting points alongside it.

What a pack installs

Depending on the pack and adapter, an installation can add:

Skills — repeatable workflows your agent discovers from your request.
Subagents — specialist executors or reviewers that receive bounded tasks.
Hooks and commands — adapter-supported automation and explicit entry points.
Seeds and configuration — repository scaffolding or pack defaults delivered through declared contracts.

The result is a set of readable files in your repository or user-scope agent directory. agentbundle installs and upgrades them; it is not a hosted agent runtime.

Step 1: install the CLI

python -m pip install agentbundle

Before any real install, add --dry-run to preview the rendered files and any preserved companions.

Claude users can also add this catalogue as a plugin marketplace for user-scope packs. core and other repo-scoped packs install with agentbundle so the workflow is shared and versioned with the project.

Step 2: choose your starting work

Install the flagship supervised work loop into a repository:

agentbundle install --pack core

Start with a feature or bug whose tests and diff you can inspect. The loop makes assumptions explicit, plans against the repository’s real contracts, implements, runs mechanical gates, and sends the diff into a cold independent review. It fixes findings until clean, stops if progress stalls, and never crosses the human merge gate for you.

Install the portable strategy, evidence, and shaping workflows you need:

agentbundle install --pack product-strategy --scope user
agentbundle install --pack desk-research --scope user
agentbundle install --pack product-engineering --scope user

You can install these independently. Together they move from a strategic question or uncertain idea through evidence and competing shapes to a decision brief a delivery team can act on. A human ratifies the bet before it crosses into implementation.

Install the supervised build spine, reviewable Terraform workflow, and release loop into the repository:

agentbundle install --pack core
agentbundle install --pack iac-terraform
agentbundle install --pack release-engineering

iac-terraform stops at a reviewable plan; it never runs terraform apply. release-engineering converges in an ephemeral environment and stops at the human production gate. Add user-scope architect and contracts when the system and interface decisions also need shaping.

Use a profile when its single-scope pack set matches your starting job:

agentbundle list-profiles
agentbundle install --profile inception

inception supports idea-to-buildable-repo work; solution-architect groups architecture, research, and contracts; full-ceremony adds the repository governance set around core.

Step 3: make the first request

Use natural language. Installed skills are discovered from the work you ask for.

Other credible first requests:

Product: “Compare the evidence for these three product bets and help me place one bounded bet.”
Infrastructure: “Design the deployment boundary, then generate a reviewable Terraform plan. Do not apply it.”
Catalogue owner: “Initialize a catalogue for our organization and show me the contracts I need to customize.”

Step 4: inspect the result and boundary

The first-value artifact should be something you can judge in minutes: a decision brief, architecture note, contract, tested diff, Terraform plan, documentation change, or release-readiness record. Check what the agent did, what remains unverified, and which decision it surfaced to a human.