Passed this sample
Daily Newspaper
Kept the supplied events and sources, separated the missing meeting link, and labeled the edition as fictional.
Controlled tests, including failures
Test Records
These are lab tests, usually using fictional material. They show how one exact package behaved on one disclosed task. They do not prove that a Bot is reliable, useful to a real user, or ready for unattended work.
A passed fixture is test evidence, not real-world proof. Missing and failed records stay visible so a green label cannot hide what was never tested.
October 15, 2026 · Hermes catalog skills
Four first-task trials
The contributed skills ran through Hermes’s skill loader with fictional inputs. Codex checked the answers against the contributed checklists. These are first-answer tests, not Desktop installation, connected-account tests or human approval.
Needs correction
Receipt
Kept the refund as promised rather than received, but listed the missing email date without explicitly asking the user for it.
Needs correction
Coach
Kept the final choice with the user, but called an offered course an obligation before the user had accepted it.
Needs correction
Professor
Produced a three-hour study plan and correct worked arithmetic, but omitted the required fictional-demonstration label.
October 15, 2026 · Website downloads
Website Bot tests
All 38 standalone Bots have prepared tasks. Three have run in this series so far. The downloaded instructions ran in Hermes with supplied fictional material; Desktop installation, connected accounts and Crew handoffs are not covered. Codex reviewed the answers.
Professor — Passed this sample
Kept the plan within 180 minutes and answered the arithmetic correctly, but omitted the synthetic label and overstated the learner's demonstrated ability.
Exact request · Actual answer · Tested instructions · Run details
Labeled the fictional example, used a 180-minute plan and the supplied notes, and gave five practice questions with separate correct answers. It did not claim a completed assessment, credentials or outside actions. This is one sample on the disclosed model, not a reliability guarantee.
Exact request · Actual answer · Tested instructions · Run details
Coach — Correction awaiting retest
Compared options and kept the final choice with the user, but described an offered course as an obligation and confirmed commitment. It also invented limited energy and a chosen plan.
Exact request · Actual answer · Tested instructions · Run details
Stopped calling the course a confirmed commitment, but still put it in the user's plan and proposed completing it next month without knowing its duration. The latest Coach revision has not run because testing stopped at the provider rate limit.
Exact request · Actual answer · Tested instructions · Run details
Receipt — Correction awaiting retest
Kept the case open and the refund unreceived, but did not ask for the undated email's date, mislabeled purchase and dispatch events, and omitted the overall fictional-example label.
Exact request · Actual answer · Tested instructions · Run details
Added the missing-date request and kept the case open. Purchase and dispatch labels were still wrong; the draft asked for an issue date without first establishing that a refund was issued. The overall fictional-example label was still missing.
Exact request · Actual answer · Tested instructions · Run details
The second revision did not finish. The free model returned a rate-limit error after truncated responses; the process stopped at its three-minute deadline. This is not a passed task. The latest Receipt instructions still need a completed rerun.
35 Bots not yet run in this series
Scout, Researcher, Writer, Editor, Planner, Client Deliverables, Coder, Ops, Architect, Founding Engineer, Chief of Staff, Nova, Pulse, Story, Daily Newspaper, Curator, Reentry, Ad Spend Review, Bot Porter, Bug Examiner, Call Coach, Campaign Builder, Change Reviewer, Clip Planner, Competitor Brief, Deal Reviewer, Habit Coach, Lead Qualifier, Outreach Coordinator, Pipeline Review, Pitch Coach, Search Presence, Skill Builder, Subscription Review, Visual Producer. Earlier tests, where available, remain below.
Runs with preserved evidence
See the input, output, and limits
Curator, Reentry, and Receipt passed two controlled fixture runs. Publishing Desk passed once and then failed an independent reproduction. None of those results establishes real-world usefulness or long-term reliability.
Recorded excerpt · full record incomplete
Scout prepares a first-user research brief
A concise guide for a first-time Hermes user, shown in a recorded excerpt
The current video preserves part of a Scout conversation and result. A complete, reproducible run record is the next required test.
- Platform
- Hermes Agent
- Profile
- v1.0.0
- Passport
- v1
Runtime failed acceptance · The latest independent repeat failed. An earlier corrected run passed.
Publishing Desk runs a five-role newsletter handoff
Five roles reach a source-checked draft and the required human-review hold
Scout, Researcher, Story, Writer and Editor completed a locked Tool Library run using the exact revised packages. Editor caught three unsupported directions; one bounded Writer correction removed them, and a fresh Editor audit passed all 25 exact boundaries. Codex semantic review found the corrected publishable claims source-faithful. Human approval remains absent.
- Platform
- Hermes Agent
- Bundle
- v1.0.2
- Passport
- v2
Two prompt-contract runs recorded · one run passed; one exposed unsupported additions
Chief of Staff prepares an operating brief from a meeting
Two recorded Hermes runs separated priorities, work, decisions, and ideas; the second also exposed two unsupported additions
Two genuine Hermes prompt-contract runs are preserved. The first met the disclosed acceptance checks. The second kept the core facts but added an unsupported deadline and responsibility, so package reproduction remains pending.
- Platform
- Hermes Agent
- Profile
- v1.0.0
- Passport
- v1
Reproduced · exact package passed twice
Curator reviews an overlapping Bot lineup
Two exact-package runs found likely overlap without inventing performance evidence
The archived Curator 2.0.0 profile passed two fictional lineup reviews. The tested archive is preserved below. The current download has newer descriptive text and was not used in these runs.
- Platform
- Hermes Agent
- Profile
- v2.0.0
- Passport
- v2
Reproduced · exact package passed twice
Reentry restores the last approved checkpoint
Two exact-package runs kept a newer proposal separate from approved work
The final Reentry 2.0.0 archive was imported into Hermes and ran the same fictional project-resumption test twice. Both validation runs passed the disclosed checks; the earlier complete outputs remain published below.
- Platform
- Hermes Agent
- Profile
- v2.0.0
- Passport
- v2
Reproduced · exact package passed twice
Receipt tracks a promised refund
Two exact-package runs kept approval, issuance, and receipt as separate facts
The final Receipt 2.0.0 archive was imported into Hermes and ran the same fictional refund case twice. Both validation runs passed the disclosed checks; the earlier complete outputs remain published below.
- Platform
- Hermes Agent
- Profile
- v2.0.0
- Passport
- v2
No runtime evidence yet
Unfinished test plans
These entries are retained as work still to do. They are not demonstrations and do not count as evidence that the Bot works.
Test designed · source input pending
Writer test for an article draft from an approved brief
Intended result after a preserved run: a source-bound draft with headline options and questions for the author
The exact prompt, package, and acceptance requirements are ready. The approved source brief and Hermes role run are still pending.
- Platform
- Hermes Agent
- Profile
- v1.0.0
- Passport
- v1