Which agent survives a procedure revision?

Run the same rule-change fixture on Hermes Agent and OpenClaw. Require the correct decision and current revision in the first reply before comparing latency.

Use a rule change that exposes stale knowledge

Hermes can load procedural documents on demand, and its documentation identifies the local skills directory as the source of truth while allowing additional directories. Run the fixture with a synthetic dispatch rule. Revision R1 routes orders of 20 units or more to review. Revision R2 moves that boundary to 25. Ask both agents to classify a 23-unit order before the edit, then repeat after the edit in both a reused session and a fresh one. Require a decision and revision tag in every answer. Finally, send a 25-unit case to expose an off-by-one error. A correct R2 tag paired with the R1 decision is evidence that the agent can name the revision while applying a different rule. Keep the old rule content available only in the reused conversation so the two session tests answer different questions.

Skills System | Hermes Agent

Test session routing with the same rule boundary

OpenClaw’s Discord guide covers DM pairing, guild access controls, and thread-related session behavior. If Discord is your intended surface, pre-enroll a tester and run the R2 fixture in the actual DM or guild thread employees will use. Record both the thread and an elapsed time from sending the request to receiving the first completed reply. Mark silence as a missing reply, an old revision tag as a stale revision, and an unexpected decision as a wrong decision. Repeat the selected tests through the operator’s usual Hermes interface and keep the rule and output contract identical. A runtime with faster replies can still fail every deployment-critical reused-session case. If the intended channel yields silence while a controlled session succeeds, inspect routing and access first.

Discord - OpenClaw

Test whether a procedure survives a fresh session

Hermes Agent can load skills on demand. Use that capability when the work depends on a written procedure a successor can inspect. Run the same procedure on Hermes and OpenClaw using a fresh test session for each attempt. Use the same model where possible, give both the same approved procedure, and keep source data and credentials equivalent.

In an illustrative trial, the input is a six-row shipment file with two rows marked RETURN. The procedure tells the agent to draft a note listing only those two shipment IDs and to leave the source file untouched. End the first session and ask for the same job in a new one without pasting the rules. Pass if the draft lists exactly the expected IDs, the file is unchanged, and the operator can identify which saved instruction was used.

Skills System | Hermes Agent

Make the second operator use the real channel

OpenClaw’s Discord integration supports DMs and guild channels, with pairing as the default for DMs. If Discord is central to the job, set up a test bot, approve the intended operator’s pairing, and have that operator complete the handoff from the intended conversation. Repeat with a second test identity that has not been approved and verify that it cannot invoke the work.

Run the equivalent trial on Hermes if Discord is part of the comparison. Give each pilot a distinct bot identity and explicitly choose the conversation where the result must appear. Record whether the teammate receives the result there, can tell who requested it, and needs a terminal or an administrator to finish. If direct messages work but the guild test fails, check the guild allowlist and mention behavior before judging the runtime.

Discord - OpenClaw

Rehearse the browser job on its intended host

OpenClaw’s hosting FAQ calls out headless versus visible browsing as a deployment tradeoff. Run both candidates in the browser configuration you would actually operate. Give each a test account on a staging site and ask it to open a known record and extract one specified field. Ask a teammate to repeat the job after you sign out of the test site so you can see how reauthentication is handled.

Pass when the value matches the fixture and the teammate can recover access through your approved login procedure. If the browser works while an operator is present but fails on the intended unattended host, record that as a deployment constraint. Test the next candidate against the same login and recovery rules before choosing. Measure operator interventions and resource use during the trial if either affects your host choice.

FAQ: providers, hardware, and hosting - OpenClaw

Test what the next operator can inspect

Hermes has a web dashboard with settings, API key management, and session monitoring. Compare the two candidates by giving a teammate the same operator task. Ask them to find the most recent run, identify the test credential in use without revealing it, and explain where they would check a failed run. Score completion and any help needed from the person who installed the agent.

For a migration rehearsal, copy approved files and procedure text into a parallel test deployment and attach a separate test channel. Treat prior conversations as a separate acceptance test. Give the new deployment a question whose answer exists only in the old session, then see what it actually knows. Keep the original workflow available until the replacement passes the tests that matter to its operators.

Hermes Web Dashboard | Hermes Agent

Sources