Compare
Hermes is one agent that grows with you. Paperclip is an org chart that any agent can be hired into. Benzene is the shape in between: cells that prove their work, an organisation founded by its own record, a heartbeat with a CEO inbox. The table is written from the public sites and repositories as of September 2026; where a product does not document something, it says so.
| Benzene | Paperclip | Hermes Agent | |
|---|---|---|---|
| Runs as | one Python process, local files, any model | one Node process with embedded PostgreSQL and a React UI | a self-hosted CLI agent on your server |
| Unit of work | a cell with a guardrail: six atoms per instruction | a hired agent woken by a heartbeat | one agent with skills and memory |
| Proves the work | ten verifier types, rollback ratchet, incrementality test, blind-spot check | approvals and audit; no verifier documented | not documented |
| Organisation | six offices founded by the record, chartered by the evolver | org chart you draw: titles, reporting lines, goals | none |
| Heartbeat | duties per office, the COO at the centre, before/after reports | DB-backed wake-up queue per agent, cron, webhook, API | cron-based automations |
| Where an instruction goes | a decision model (Jev) picks the office and the cell as a typed choice with a confidence; the COO plans when it is unsure | issues assigned to an agent on the org chart | one agent takes everything |
| Approvals | CEO inbox, operator chain, features only | board approvals, execution policies | none |
| Budgets | per-task ceiling, treasury ROI verdict; monthly caps on the roadmap | monthly per agent, 80% warning, hard stop | not documented |
| Crash recovery | execution log, snapshots, recovery on start | orphaned run recovery | not documented |
| Where the guardrails come from | research: one subagent per office documents the competition, the standard's bars become checks proven on the small model | written by hand per agent | not documented |
| Publishing | an outbox per channel: held until the office is healed and access is granted, a few items per beat under a budget | agents act through their tools | chat platforms |
| Skills | portable SKILL.md, developed from verified executions, guardrails proven by the scout | Skill Studio, shared org-wide | 40+ built in, auto-created, agentskills.io |
| Growth of the agents themselves | heal, evolve on measured gain, divide, die, regenerate | not documented | persistent memory, new skills |
| Evaluation of the product | Product Hunt and Shark Tank panels, launch, operations and build reviews | not documented | not documented |
| Bring your own agent | any model; external agents on the roadmap | any runtime that can receive a heartbeat | it is the agent |
| Chat platforms | none; the terminal and mail | not the focus | Telegram, Discord, Slack, WhatsApp, Signal |
| Dashboard | a graph view and the inbox as files | web UI | CLI |
Paperclip: paperclip.ing and github.com/paperclipai/paperclip. Hermes: hermes-agent.org. Both are open source and self-hosted, as is Benzene. If a row misstates a product, open an issue and it will be corrected.
You want one agent on your server that remembers you, reaches you on chat, and grows skills as it goes. You check its work yourself.
You have a product with checks of its own (a compiler, tests, documents with structure) and you want agents whose work must pass them, an organisation that grows around the product, and a heartbeat you approve from an inbox.
You already have agents on several runtimes and want an org chart, budgets and board approvals over all of them, with a web dashboard.
Benzene