Where Did Your Agent Stop Making Progress?

Send a harmless probe through both the local agent and its messaging gateway. Use one trace ID and separate budgets for ingress, execution, and delivery to find the first late event.

Run the same probe through two entry points

Hermes Agent’s quickstart verifies a real chat before adding a bot or shared assistant. Use that sequence as a diagnostic control. Run a probe locally that asks for a fixed, harmless response containing a trace label, then send an equivalent probe through the affected gateway with a fresh label. Keep the model configuration and prompt the same. Record submission, runtime acceptance, result creation, and reply receipt. If the local probe succeeds while the gateway probe has no acceptance event, inspect the channel binding and gateway route first. If both probes reach the runtime but neither produces a result, inspect the provider response and execution logs. Repeat the pair after one configuration change. A visible reply bearing the correct trace label lets you distinguish a recovered path from a delayed response to an earlier attempt.

Hermes Agent Quickstart | Hermes Agent

Measure progress across an execution pause

LangGraph documents durable execution and human oversight of agent state. For an agent with that kind of checkpointing, build a staging probe that waits at a known checkpoint before returning a read-only result. Record the checkpoint ID and emit a separate heartbeat for the worker. Pause the test, resume from the recorded checkpoint, and confirm that the next expected state appears under the original run ID. A live heartbeat with a stationary checkpoint tells you the process is responsive while the workflow still lacks progress. An advancing checkpoint with no reply receipt shifts the investigation downstream to delivery. For example, give the read-only step an illustrative 30-second budget after resume and check the last two checkpoint IDs when it expires. This gives the operator a concrete place to inspect before taking any recovery action.

GitHub - langchain-ai/langgraph: Build resilient agents. · GitHub

Find the transition that never happened

Look up the run by ID and compare the last recorded transition with the next event you expect. A queued run needs a start record and a queue-age limit. A tool call needs a result or a timeout. A completed artifact needs a delivery record. For a deliberate approval pause, check that the approver, deadline, and resume identifier were saved together before the worker released the run.

Probe the runtime independently before replaying the task. In OpenClaw, a fresh Quick start with usable provider access verifies that access with a real completion and opens the web dashboard. Use a harmless message in a known-good session as a separate health check. If it replies while the affected run stays queued, check admission slots and worker leases before touching the run's checkpoint.

Getting started - OpenClaw

Rehearse approval after a worker restart

In an illustrative test, run R42 produces a draft at 09:00 for reviewer Maya, with approval due at 09:15. Save the draft location, run ID, reviewer ID, deadline, and next step together. Restart the worker at 09:05, then send the same approval event twice at 09:10. Pass if the saved draft resumes once and the second event records that the approval was already applied. Repeat with no reply and verify that the draft remains unpublished and the run expires.

When the trigger is a Hermes interactive install waiting for operator input, configure the unattended install path separately. Hermes documents flags that skip stages requiring input. Supply any required configuration before running a live chat check. An installer exit alone leaves provider and gateway readiness unverified.

Install Hermes Agent: Official Guide | Nous Research

Test a tool timeout without repeating a write

Give every external call its own attempt ID and deadline. When a run stops after dispatch, look for the matching result or a status record at the downstream service. A read can usually be retried within a bounded budget. For a write with an unknown outcome, check whether the downstream effect already occurred before replaying it. Use a stable operation ID when the destination supports deduplication, and route ambiguous writes to review.

For a Hermes smoke test, start its documented terminal interface and check for the selected model in the banner. Send a basic prompt, then ask it to read a known local fixture and follow up about the result. If plain chat works but the fixture read produces no result, investigate tool dispatch, authorization, and the timeout path before changing providers.

Hermes Agent Quickstart | Hermes Agent

Recover delivery from the saved result

If a completed artifact exists, trace the send attempt separately from generation. Inspect the destination, outbound request status, and any message ID returned by the channel. In the Claude Code Slack channel project, the troubleshooting guide identifies missing bot membership as a reason a bot may never see a test message. Check membership and send a harmless probe in the same channel to establish whether inbound events are reaching the bot.

When inbound events arrive and an artifact exists, test outbound delivery with the stored result and destination. Pass if the expected channel receives one message carrying the original run ID and a retry produces no duplicate. If the platform confirms a send but a user cannot see it, inspect the thread and the user's permissions before generating another result.

GitHub - jeremylongshore/claude-code-slack-channel: Slack-native governance substrate for Claude Code and peer agents: every tool call passes a tiered policy engine and lands in a hash-chained, Ed25519-signed audit journal you can verify offline. The governance kernel the companion agent-governance-plane (AGP) builds on. · GitHub

Sources