few days ago I installed @NousResearch's Hermes agent on my mac mini, alongside my existing @openclaw (Oz) bot that's been running for 6 weeks.
two AI agents. one machine. 16GB RAM. same Telegram chat. here's the journey so far.
~/ the idea:
what if a second agent with a different architecture found things the first one missed?
setup took about a day. cloned the repo, created a Python 3.11 venv, configured GLM-5 as the primary model with Kimi K2.5 as fallback. wrote SOUL.md - hermes' system prompt defining its personality, tools, and file paths. configured 11 Telegram toolsets: web, terminal, file, skills, memory, todo, session_search, code_execution, delegation, and cronjob. set up launchd to keep it alive on boot.
then I set up the research pipeline. Hermes runs autonomous research sessions 1 hour after Oz on the same topics. both write findings to separate research files and deliver bullet-point summaries to Telegram. I compare their output side by side every evening.
benchmarked them head to head on 8 identical queries. Hermes scored 34/40, Oz scored 36.5/40.
- Hermes was dramatically faster (5 seconds vs 60 seconds) and better at structured data retrieval, live prices, protocol stats, CoinGecko pulls.
- Oz was stronger at accumulated context, writing in my voice, and connecting new findings to past research.
early results are promising. in its first 48 hours of autonomous research, Hermes surfaced the Meta/Moltbook that happened today.
~/ about bugs:
mostly on my end, some on Hermes'.
- the web search tool needed a Firecrawl key I hadn't configured, so I rewired it to DuckDuckGo CLI.
- the API endpoint and model naming conventions were different from what I expected, took trial and error to get right.
- reasoning_effort doesn't work on this provider.
- sessions bloat after 30+ turns because reasoning tokens accumulate, so I built automatic compression, daily resets, and a weekly cleanup script.
- the best one: Hermes couldn't find its own cron research because the system prompt didn't specify where those files live, it concluded its own findings were hallucinated instead of admitting it didn't know the path.
every bug was fixable, and each one taught me something about how agent infrastructure actually works.
~/ here's why I think Hermes could genuinely surpass OpenClaw over time:
it runs natively, no Docker overhead, no sandbox boot, direct filesystem access. it has native tool schemas built in (terminal, file operations, code execution) rather than everything running through a gateway. and the cron system, once patched, is clean and lightweight.
Hermes is few days old with few days of memory. Oz has 6 weeks. if Hermes' knowledge compounds the same way Oz's did, with the speed advantage on top, it could become the primary agent.
for now I'm running both in parallel. same topics, different timing, comparing output daily. I'll report back in a few weeks on which one produces more actionable research, which one catches more breaking news first, and whether Hermes' speed advantage outweighs Oz's depth.
two agents on one mac mini for $0/month in infrastructure. the experiment continues.