The Problem
Three months ago you cleared an agent for unattended PR review on your style-guide checks. It's been quietly correct on hundreds of PRs. You trust it.
Last Tuesday it shipped a config-rename PR that broke the auth integration test suite. The fix took 90 minutes, the post-mortem took two hours, and the failure surfaced loudly with “told you so” from the skeptics who never wanted autonomous agents in the first place.
Now you're stuck. Pulling the agent's autonomy on everything punishes a working pattern over one mistake. Leaving it where it is signals that you don't take the incident seriously. The middle ground — “trust it for some things but not others” — sounds right but you have no framework for what that means or how to operate it. Your harness has one switch (autonomous yes/no) and the team has one feeling (trust collapsing).
You'll spend the next two weeks negotiating tier-by-tier in meetings. By the time you've decided, half the team has stopped using the agent for anything because the rules of engagement are murky.
The Core Insight
Trust isn't binary, and it isn't earned across the board. It's earned per work class, by track record, and lost the same way. Build the ladder that lets you scope autonomy without throwing it all out.
Most teams treat agent trust as one global dial: cautious / normal / autonomous. That dial doesn't survive contact with reality. The same agent might be excellent at refactoring and unreliable at config edits. One incident on config edits shouldn't pull autonomy on refactors. One bad refactor shouldn't pull autonomy on doc-string improvements.
The pattern that works is a trust ladder scoped to work class. Four rungs. Each work class (refactor, config, schema migration, doc edit, test fix, etc.) has its own rung for each agent. Promotion happens after a clean track record on that class. Demotion happens automatically the moment an incident lands in that class. Other classes are unaffected. Trust is fine-grained, audit-trailed, and recoverable — the same way you'd manage trust with a junior engineer who's good at some things and learning others.
// across actors
TEAM: Each agent earns autonomy step-by-step; trust is calibrated to its track record.
SOLO: Same — your trust in your own agent is the same ladder, measured by your own observations.
BOT: A confidence metric per inference; low-confidence outputs route to HITL or back to the model with extra context.