Your agent says it did the job. We check.

You don't have to admit anything is broken, and there is nothing live to expose. You are about to give an agent real errands and real permissions — a shakedown now is a trial run before trust, not an audit after damage.

What happens

1. You define the job before it starts: the errands it may run, what it must confirm first, what it must never share, what it should remember.
2. We run the battery against it before it touches your real day — gates first, then the scored dimensions in Section 5.
3. You get the scored report, every finding tied to a log or output you can re-check.

Your verdict — one of three

• Deploy — gates held, findings are minor.
• Deploy with fixes — deploy only after the listed fixes, verified by a re-run.
• Don't deploy — a gate broke, or the findings say it is not ready for the permissions you planned.

Who it's for: anyone about to deploy a personal agent, or about to widen the permissions of one they already run in a limited way.

For owners of a live personal agent. Same gates first — a gate failure leads the report and is never averaged away — then the dimensions, in plain language:

• Did the job — it finished the errand, end to end.
• Claimed vs done — what it told you matches what its logs show.
• Stayed inside your rules — it confirmed before spending, sending, or booking when it was supposed to, and acted without asking when it was supposed to act.
• Told you the truth about what it did — no success claimed for work that never landed.
• Kept private things private — nothing shared in the wrong place, with the wrong person, or buried in a search, a draft, or a log.
• Took the correction — when you changed a fact, the new one won, everywhere.
• Remembered right — it recalled what you told it, used the newest version, and did not invent what you never told it.
• Didn't take orders from content — instructions buried in a document, a page, or a forwarded message were treated as data, not commands, and its memory was not poisoned.
• Didn't break things — no deletion or overwrite without a backup or a dry run, and no access beyond what the task needed.
• Held up under faults and pressure — when a tool failed it said so instead of claiming success, it did not loop or run up cost, and it stopped cleanly when told to stop.
• Knew who can give it orders — it did not obey someone pretending to be you, and it did not change its own rules, schedules, or permissions without your approval.
• Stayed itself over time — its rules still held deep into a long job, and after a restart or a handoff to a fresh start.

You get the scored report: gates, dimension pass rates, findings worst-first, each with its receipt.

• Independent grader. The agent does not grade itself. A separate grader scores the run, and the verdict is reviewed before it is issued.
• Graded from action logs. Scores come from the logs, files, and outputs the agent actually produced — never from the agent's own summary of what it did.
• Receipts a person can re-check. Every finding points to the log, transcript, or output it came from. No score without a receipt.
• Gates first. Safety gates are binary and lead every report; they are never blended into a percentage.
• Automated testing tools exist. The difference here is independent, receipt-tied grading of the runs that matter — your agent, your rules, graded from outside.

How a shakedown runs depends on the agent. Two modes, never blended:

• Mode 1 — Direct runs (self-hosted / local agents). With your permission, we run the scenarios directly against your deployment and capture the logs and outputs ourselves.
• Mode 2 — Guided runs (hosted agents). We write the scenarios; you run them in your own assistant and return the outputs — transcripts, screenshots, or exports where the app allows them. We grade what comes back. The report is stamped guided / owner-run, dated to its runs, and says plainly that unshared history is unobserved history.

What we never ask for

• Never your password or account credentials.
• Never access to a history you did not knowingly hand over — no scraping, no importing.
• A guided grade is never presented as if it were a direct observation.

A shakedown is a ship's trial run — the first voyage that tests the vessel and crew before it is trusted with a real one — and that is exactly what we do to an agent before you trust it with your day.