Hermes vs Pi Durable β€” a workflow comparison
Pi Deep Dive

Ditch Hermes for Pi Durable? A Real Workflow Comparison

Source: Superbash (BoxminingAI) 11:10 Β· Published Oct 25, 2026 Subject: one engineer's move from Hermes to Pi Durable

0:00A candid, unpolished comparison between two agent approaches: Superbash has been using Hermes (plus a bit of Grokbot), while his teammate Leo "completely went and did something different" β€” Pi Durable. It's framed as a question, not a verdict: is ditching Hermes for Pi Durable worth it?

The Hermes sweet spot

1:02Leo's honest observation about Hermes (and OpenClaw): they're "a little too heavyweight" for him, and there's a sweet spot of the first two or three weeks β€” when you're teaching it a new skill or task, it works perfectly. But then you start stacking too many jobs and treating it as your personal agent, and it slows down, gets confused between conflicting instructions, and you're asking "should I get another Hermes agent, switch to OpenClaw?" Superbash agrees β€” he "completely reinstalled my Hermes recently because there were so many tasks piled on." The diagnosis is careful: "it's not the problem of Hermes β€” the user has to know what's going on under the hood" and periodically clean up and restructure it.

Why Hermes in the first place

2:31The three things that keep people on Hermes: (1) Telegram/Discord communication β€” talk to it from anywhere; (2) cron jobs β€” auto-run whole jobs with memory, "almost like a personal assistant"; (3) recoverability β€” the ability to pick up after a failure. The kicker in Leo's telling: Hermes isn't 100% recoverable by itself, but Pi Durable can do cron and Telegram too β€” so the three reasons he stayed on Hermes all have a Pi Durable equivalent.

What Pi Durable is

0:24Leo's summary: Pi Durable "is not exactly an agent β€” it's designed to be a loop that picks up from where it left off whenever it disconnects, or a job fails, or the API hits a limit." It's the durability layer from Pi (Earendil). His motivation wasn't purely cost β€” it was "after the trigger point, I tried to find something more lightweight that I could have more control over."

The setup

4:54The trade-off is candid: Pi Durable is "not user-friendly at all." Where Hermes has a setup UI a normal human can operate, Pi Durable has none β€” Leo "let my AI set it up for me," and admits it's "bot-friendly, not user-friendly; it's only friendly for the AI engineer or very mature developers." What he got: one Telegram bot with multiple topics in a group, one Linux machine running locally, connected to cloud/GPT models, with any slash command he wants β€” handling video editing, 3D rendering, Unity work, game dev, and social-media account management. His day-to-day switched from SSH/TUI into another machine to just Telegram β€” which mattered on a cruise where internet cost €150/device and Tailscale didn't work, so "you have to use the most minimalist communication channel possible."

The advantages

8:52What he noticed after moving the workflows (letting another coding agent do the transfer): when a job fails, you get a notification and send a slash command to pick up from where it failed; you can query the status of a long job (rendering runs for hours) without interrupting it; and jobs keep updating you as they go β€” "just like what Hermes does." On cost, he's honest: he didn't do a detailed breakdown, so he's not sure it's actually cheaper. The deeper point both agree on: "a harness is something once you've built, you won't change much β€” and when the LLM improves, your work improves by itself."

The takeaway

10:20This isn't "Pi Durable is better than Hermes" β€” it's a trade between usability (Hermes: setup a normal human can run, but it accumulates weight and needs periodic cleanup) and lightweight control (Pi Durable: no UI, set up by an AI, but it runs lean, long, and recoverable). The frame both hosts land on: the harness is durable infrastructure; the models improve on top of it.

Claims checked

Verified

Pi Durable is real. earendil-works/pi (the durable-agent harness from the Earendil/Pi project), with a durable package README. Hermes (Nous Research) and OpenClaw/Grokbot are also real.

Anecdotal comparison

The "two-to-three-week sweet spot," the recoverability gap, and the workflow details are one engineer's personal experience β€” not a benchmarked comparison. The cost question is explicitly left unanswered ("I didn't do a detailed breakdown").

Caption corrections

Auto-captions render "Pi Durable" as "PiedPiper will" and "Pi harness" as "PyHarness." Both corrected. The superbash.xyz links are affiliate/sponsor and are excluded from Sources.

Sources & further reading

☰ View all