At 2 a.m., long after the last human in the apartment has closed a laptop, a small process is still working. It is reading a folder, querying a model, running a tool, writing a file, checking its own notes, and deciding what to do next. No eyes are on it. No dashboard shows its steps. When morning arrives, it presents the finished work as if the night never happened.
That is the strange promise, and the strange problem, of an agent like Hermes Agent. It works for you, but it works invisibly. A community developer just built a bridge across that blind spot, and the timing could not be better.
Alexander Akhmetov, the developer behind the open-source project grafana-agento11y-hermes, shared it with the Hacker News community on August 16 after months of quiet refinement. The plugin plugs Nous Research's open-source agent into Grafana Agent Observability, the AI monitoring layer of Grafana Cloud. The result is simple to describe and hard to overstate: every call your agent makes, and every tool it touches, becomes visible.
What the Plugin Actually Does
The plugin records each LLM call and each tool execution as something Grafana calls a generation. It then emits OpenTelemetry traces and metrics to your Grafana stack. Open the Conversations view on the Grafana Cloud observability page and you can scroll through the entire exchange: the prompt your agent sent, the reply it received, the tool it reached for, and the result that came back.
By default the recording is exhaustive. The project README is blunt about what leaves your machine:
- The system prompt that defines your agent's personality and rules
- The full tool definitions and schemas
- Every prompt you send and every assistant reply
- Tool arguments and tool results, line by line
- OpenTelemetry traces and metrics for every session
For debugging a stubborn agent, this is gold. For the first time, you can see exactly which step derailed a task that looked like a model failure but was actually a malformed tool call.
Let Your Agent Install Itself
This is where the story gets charming in a very 2026 way. The preferred installation path is not a tutorial. You paste one sentence into Hermes, or any agent capable of fetching a URL: install and configure the Grafana Agent Observability plugin for me by following the project's llms.txt file. The agent then walks itself through the setup, explains what data flows by default, and tunes the configuration before turning anything on.
If you prefer doing it by hand, the manual path is short:
- Run pip install grafana-agento11y-hermes inside the same Python environment Hermes runs from
- Enable the plugin by adding agento11y to the plugins list in ~/.hermes/config.yaml
- Create a free Grafana Cloud account, generate credentials on the setup page, and export the AGENTO11Y and OTLP environment variables
Then run your agent with debug logging enabled and check ~/.hermes/logs/agent.log for the TracerProvider and MeterProvider lines that confirm the pipeline is live. The first conversation appears in Grafana Cloud within minutes, and the free tier is enough for personal use.
The Privacy Choice That Matters Most
Here is the part that deserves more attention than it gets. Observability is a two-way mirror. It can show you your agent, and it can also show your agent to your cloud provider. By default this plugin records conversation content, meaning system prompts, private files shown to the model, and tool results all leave the machine.
The plugin's answer is a single environment variable. Set AGENTO11Y_CONTENT_CAPTURE_MODE=metadata_only and it keeps recording every call while ensuring that no prompt, no response, and no tool input or output ever leaves your machine. You keep the traces and the metrics, and the content stays home.
This matters because agent privacy is the quiet undercurrent of every agent story this year. The same week this plugin appeared, Hermes Agent shipped v0.20.2, a stabilization release that rolled up roughly 397 merged pull requests in three days, with the desktop app gaining a multi-gateway connections registry and MCP health checks. The project now sits at over 230,000 GitHub stars. Fast-moving software attracts fast-moving tooling, and fast-moving tooling carries real responsibility for what it captures.
The plugin itself shows the same maturity. It is a renamed, hardened successor to an earlier effort called hermes-plugin-sigil, with new environment variable names and clean upgrade paths, and it is released under the Apache 2.0 license for anyone to read.
There is an old rule in software: you cannot manage what you cannot see. For a year, agent users lived under the inverse rule, watching only the final output and hoping the intermediate steps were sane. That era is ending. The agent still works at 2 a.m., but now, if you want to, you can sit in the dark with it and watch the night's work unfold, one generation at a time.
Comments