Can Hermes Agent or Claude Code Clear a Backlog?

Evaluate Hermes Agent for recurring bot work and Claude Code for repository execution. Compare each deployment’s measured drain capacity against the same backlog and recovery deadline.

Measure service rate with distinct recurring jobs

The Hermes Agent quickstart treats team, bot, and always-on work as a setup goal and calls for verifying a working chat. Extend that smoke test with a short batch of synthetic recurring jobs, each with a unique ID and a verifiable result. For example, ask the deployment to report the count of rows in the same small fixture under six distinct job IDs. Measure elapsed time from the first accepted job to the last distinct correct receipt, including model or tool waits. Record duplicate receipts separately and count only correct, unique completions in the throughput numerator. A bot that responds quickly to the first request can still fall behind under sustained arrivals. Repeat with the tool mix used in production, since a simple receipt can overstate throughput for jobs that read files or call other tools.

Hermes Agent Quickstart | Hermes Agent

Account for what survives the execution process

Anthropic’s hosting guide describes the Agent SDK worker as a long-lived Claude CLI process with a working directory and session files. For repository work, measure recovery with an authorized test worker and a synthetic queue whose job IDs and expected outputs you retain outside the worker. Stop intake to the worker during the drill while continuing to add synthetic rows to the queue, then resume processing and count distinct accepted outputs. Include any time spent reattaching to state or rebuilding the checkout in the outage measurement. If an ID disappears from the queue without a checked output, count it as missing even if the queue length is now zero. The capacity calculation assumes every pending ID survives. Verify that assumption with the job records before using projected drain time to set an overnight recovery target.

Hosting the Agent SDK - Claude Code Docs

Use Hermes for a timed delivery

Hermes Agent documents built-in cron, so it is a candidate when a job starts at a set time and the result needs to reach a chat. Start with one read-only duty. Choose a provider through the quickstart, confirm an ordinary chat works, and then set up a single scheduled run against a fixed test input. Ask for a timestamp, the input identifier, and a short result.

Watch for one message in the intended conversation within the delivery window you set. A duplicate, stale input identifier, or missing message is a failed run. If delivery fails, test a direct chat again and check whether the scheduled job ran. A successful direct chat alongside a missing scheduled result narrows the investigation to scheduling or delivery.

Hermes Agent Docs (Official) | Nous Research

Use Claude Code for a patch with a test

Claude Code is a good fit when the deliverable is a repository change someone can review and reproduce. Its documented CLI can edit files and run commands. Start it in a disposable checkout at a recorded revision. Give it a narrow failing case, specify which files it may change, and name the tests to run before it hands back the diff.

In an illustrative trial, a report formatter prints a blank amount when a fixture contains zero cents. Give Claude Code that fixture and require the formatter to display 0.00 while leaving handling for missing values unchanged. Accept the result only if both cases pass, the diff stays within scope, and a reviewer can repeat the checks from a clean checkout.

Overview - Claude Code Docs

Close the session during the drill

Claude Code channels push events into an open session. If you plan to send repository work from chat, use a test channel and a harmless request for the current revision. Send one request while the session is open and one after it closes. Reopen the session and send a third with a new identifier. Record which requests are acknowledged and whether a human can tell which ones still need handling.

For a resident chat duty, run a comparable drill with Hermes by restarting its test process between two scheduled runs. Check for correct delivery after restart and test recall in a fresh conversation using a harmless fact entered earlier. If delivery resumes but recall fails, verify which state your deployment persists before assigning work that depends on past decisions.

Push events into a running session with channels - Claude Code Docs

Hand off a ticket and require an acknowledgment

When both products are useful, pass an explicit ticket from Hermes intake to a Claude Code workspace. Include a request ID, base revision, failing case, permitted paths, and an acceptance check. Return the diff and test results to the reviewer. A recorded acknowledgment distinguishes a ticket that reached the coding workspace from one that still needs an operator to pick it up.

If the coding side uses the Claude Agent SDK across multiple hosts, Anthropic documents a SessionStore adapter that mirrors transcripts to a backend you choose. Test resumption from a matching working directory and verify the expected context appears. Provision the checkout separately and confirm its revision. The handoff passes when a replacement worker can identify the correct task and reproduce the acceptance check.

Persist sessions to external storage - Claude Code Docs

Sources