OpenClaw and Hermes Agent: A Practical Evaluation
Table of Contents
Return to the Practical AI Workflow Course
Choose a specialized agent from the task and its access needs. A worker reviewing one file differs from a persistent system receiving messages, storing memory, and running scheduled actions. This optional lesson uses OpenClaw and Hermes Agent as examples, then gives you a pre-install evaluation worksheet.
Learning outcome: you will compare agent access and triggers, then record a limited pilot decision.
Key Takeaways
- Role worker: a bounded agent assigned one question or artifact.
- Persistent agent: a service or session retaining state or responding to later triggers.
- Trigger: an event, message, or schedule starting work.
- Access inventory: a list of tools, accounts, files, and write scopes the agent receives.
- Pilot decision: reject, run a limited test, or approve a defined use.
Before You Begin
Prerequisites: delegation , goal loops , and a fictional use case. Set aside 35 minutes. Difficulty is intermediate. This lesson does not require installing a persistent agent.
Use case: a workshop coordinator wants an assistant to read a schedule and prepare a reminder draft. The request does not authorize sending the reminder. Your evaluation should preserve this boundary.
Compare the Roles
| Approach | Typical task shape | First question |
|---|---|---|
| One chat | Draft one reminder | Is the schedule present? |
| Read-only worker | Check schedule facts independently | What source should it read? |
| Coding agent | Update a local report program | Which files and tests apply? |
| Persistent platform | React to messages or schedules over time | What triggers, tools, and stored state exist? |
OpenClaw is an open-source agent platform with a gateway, messaging channels, tools, memory, MCP connections, and subagents. Its overview , MCP connection guide , and subagent guide describe those components. Review the active channel, tool, and sandbox settings before a pilot.
Hermes Agent is Nous Research’s agent application. The Hermes model family is separate. Its
documentation
,
delegation guide
, and
configuration guide
cover CLI, messaging, tools, memory, and delegated tasks. Its
memory guide
describes local memory stores and write approval. The documented default for memory.write_approval and skills.write_approval is false. Turn approval on or disable the relevant store for a pilot requiring reviewed writes. Check the running configuration before relying on either control.
| Evaluation point | OpenClaw | Hermes Agent |
|---|---|---|
| Entry point | Gateway and connected channels | CLI or configured messaging interface |
| Delegated work | Subagents with platform-specific session and tool rules | Delegated tasks with platform-specific tool rules |
| Persistent state | Inspect enabled memory and session storage | Inspect memory and profile settings, plus skill writes |
| Recurring work | Inspect configured triggers and schedules | Inspect cron jobs and gateway state |
| First pilot | One manual request through a selected channel | One manual CLI or messaging request |
This table describes review points, not equivalent security defaults. Record the installed version and active configuration before granting real data or accounts.
Names do not define a common permission model. A “subagent” in one product might receive different files, tools, and history from a worker in another. Check the installed version’s documentation and active settings instead of transferring assumptions across tools.
Fill an Access Worksheet
Write one row for each capability before installing or connecting a system. Use the actual settings if you have a product available. Use “unknown” rather than guessing.
| Question | Workshop pilot answer |
|---|---|
| Task | Draft a reminder from a synthetic schedule |
| Input | One approved schedule file |
| Files | Read the schedule file, no write access |
| External tools | None for the first pilot |
| Account access | No email or calendar account |
| Trigger | Manual request only |
| Memory and skill writes | Record active settings, disable stores or require write approval during the pilot |
| Logs | Record storage path, retention rule, prompt, source revision, and draft |
| Review | Human checks date, time, place, and recipient list |
| Rollback | Discard draft and disable the pilot |
Data flow is part of the worksheet. Name where input leaves your computer, which provider processes it, and where transcripts or memory persist. If a tool connects to email, document whether it reads, drafts, or sends. Those are distinct scopes.
Make a Pilot Decision
Reject: the agent needs broad mailbox write access for a draft-only task, or its trigger and logs are unknown. No installation is needed to reject a poor fit.
Pilot: restrict input to the synthetic schedule, use a manual trigger, disable sending, and inspect one draft. The expected result is a correct 7 p.m. time and no sent message. Record your own observation if you run the pilot.
Approve a defined use: only after the pilot meets its checks and the owner accepts the exact access, trigger, retention, and review steps. Approval for drafting does not approve sending or later automation.
Diagnose a Failure
Suppose a reminder draft uses 6 p.m. Check whether the agent read the current schedule and whether stored memory or an old message supplied the stale time. Replace the stale input, rerun one manual pilot, and keep sending disabled. A correct second draft does not erase the first failure from the record.
Suppose an MCP connection appears in settings but no tool is reachable. Return to the MCP connection states . A saved server entry is not evidence of a successful read.
Practice and Check
Your artifact: a completed access worksheet and one decision, Reject, Pilot, or Approve a defined use. State one trigger, one permission boundary, one data-flow question, and one way to stop the agent.
Completion check: another reader should know which action the agent takes, what it reads, what it stores, when it runs, who checks the result, and how to disable it. If any answer is unknown, choose Reject or a narrower Pilot.
Check Your Understanding
Question: Does a draft-only reminder task need permission to send mail? Answer: No. Keep sending outside the pilot scope.
Question: Which Hermes settings need inspection before a write-controlled pilot? Answer: Check active memory and skill write settings, including their approval controls.
Next Steps
Continue with the AI Workflow Capstone . For a broader runtime comparison, use the site’s agent runtime guide .
Course navigation: Previous: Bounded Loops , Course outline .
Related Posts
