/install

// one checkout, one bounded experiment, one fresh verdict

// what you're installing

AgentOps runs one bounded experiment from fixed intent to an independently judged, content-addressed verdict. The skills run inside your coding agent. The optional ao CLI supplies deterministic checks, inspection, and skill linking. Source: github.com/boshu2/agentops.

// core loop Plan → Implement → fresh Validate → durable verdict → report and stop.

Inspect AgentOps ↗Read the 12 factors →

// install

Use one canonical checkout or copy the skills directly.

Canonical checkout (recommended)

One source checkout. The ao CLI links its skills into every detected runtime.

bash

brew tap boshu2/agentops https://github.com/boshu2/homebrew-agentops
brew install agentops
git clone https://github.com/boshu2/agentops.git ~/.local/share/agentops
cd ~/.local/share/agentops
ao skills link

Copy the skills without ao

Useful when you want editable copies and do not need the optional CLI.

bash

npx skills@latest add boshu2/agentops --all -g

// what the loop records

The current experiment and the optional knowledge layer have different jobs.

surface	job
Beads or caller intent	the work, acceptance, decisions, and evidence references
.agents/ao/intents/sha256/	the exact resolved intent bytes used by the run
.agents/ao/verdicts/sha256/	the content-addressed PASS, FAIL, or NOT_PROVEN verdict
reviewed LLM wiki	optional reusable knowledge with links back to its sources

Beads preserve the work trail. A reviewed LLM wiki preserves reusable knowledge and its sources. The wiki sits outside the binding loop and cannot change the verdict for current work.

// blast radius No daemon, scheduler, hosted control plane, retry controller, or delivery engine is required. AgentOps reports the verdict and stops. Your repository keeps its own Git, merge, release, and rollback policy.

// update and remove

Links make both operations explicit.

bash

# update
cd ~/.local/share/agentops
git pull --ff-only
ao skills link

# preview removal, then unlink only AgentOps-owned links
ao skills unlink --dry-run
ao skills unlink

Unlinking leaves foreign skills, real directories, the AgentOps checkout, and project evidence untouched. Full day-two instructions live in the upstream install guide.

// first run

Give the loop one small, testable intent.

In your agent chat, invoke /rpi with a small change and its acceptance check. Plan fixes the intent. Implement builds once. A fresh Validate context checks the exact subject and writes PASS, FAIL, or NOT_PROVEN. Then the run stops.

text

/rpi Add one focused behavior with a failing test first.

Plan → Implement → fresh Validate → durable verdict → report and stop

Same proof boundary every run.

Install AgentOps, run one experiment, inspect the verdict.

Read Factor VII →