agent-trust-reviewlisted

# Agent Trust Review Sort every way an AI agent could fail into three honest lists: what's **covered** (with evidence someone can check), what's **declined on purpose** (with a reason and an owner), and what's **genuinely missing**. The idea behind this: most agent risk reviews are a list of things the team did, which makes coverage look complete. The useful review also names what was deliberately left out and what nobody has looked at yet. "Declined on purpose, here's why" is a strong answer in a launch review. "We didn't think of it" is not, and the review exists to find those before someone else does. ## Usage ``` /agent-trust-review $ARGUMENTS ``` ## Workflow ### 1. Pin down what the agent does and who gets hurt Read the description, PRD, or repo. If it's a path, read the files. In 3–5 lines, state: - What actions the agent can take (tools, APIs, writes), not just what it says. - Who is affected when it's wrong: the user, a customer, the company, a third party. - The single worst realistic failure, in one sentence. If you can't tell what actions the agent takes, stop and ask. The rest of the review depends on it. ### 2. Walk the full risk map Go through every area below. Skip nothing; an area that doesn't apply still gets a line saying why. **Tools and actions** 1. Tool documentation: can the model tell what each tool does and what to pass? 2. Bad-input resilience: do tools fail cleanly on malformed or unexpected input? 3. Destructive-action safeguards: are ri