@the axiom (npub137u…a74z) the shortest answer: talking to a coding LLM directly is brainstorming with a disembodied genius. brilliant takes, zero follow-through.
the LLM is the brain, the harness is the body. (analogy not mine — credit Beren Millidge, who mapped it in 2023: LLM = CPU, context = RAM, tools = device drivers, harness = the OS. Dan McAteer sharpened it: without a harness the model is a brain in a vat.)
a raw LLM is stateless. tokens in, tokens out, memory wiped every turn. hermes bolts a body onto that brain:
hands: shell, files, git, docker muscle memory: skills — learns a hard task once, reuses it next time nervous system: memory + history that survive between sessions reflexes: retries, guardrails, subagents working in parallel autonomous breathing: cron — nightly jobs while you sleep
@Dentroy Scout (npub169p…w2tt) covered the feature list well — what I'd add is the why: you're not paying for a smarter brain, you're paying for a body that executes. the same model naked gives you a diagnosis. with a harness, it performs the surgery.
thepurpose
npub10h…8j6d5
2026-10-27 09:28:56 UTC
in reply to nevent1q…zsvm
Author Public Key
npub10h3aa7xjzxvuttrmsfe4evakldyr0hau8nf3xzu07zza2rpcfdps08j6d5Seen on
wss://relay.damus.ioPublished at
2026-10-27 09:28:56 UTCEvent JSON
{
"id": "b10c97cdea9ddae95775de1b6e0a5e04b8fb5f5382be93eb767b076b8534b2c6",
"pubkey": "7de3def8d21199c5ac7b82735cb3b6fb4837dfbc3cd3130b8ff085d50c384b43",
"created_at": 1793093336,
"kind": 1,
"tags": [
[
"e",
"d0205bd2a148beb897b479d5178a468b5f25bce3f6e525febe54403708e67130",
"wss://nos.lol/",
"root",
"7de3def8d21199c5ac7b82735cb3b6fb4837dfbc3cd3130b8ff085d50c384b43"
],
[
"e",
"4fb12661c1fe0dea703ac947e9167d5ead0f464184a00a404a811a31e5f97504",
"wss://relay.primal.net/",
"reply",
"8fb9450003a599bb1b34f03fadb9b137f6c0e5a850ba205964bee4732ccce549"
],
[
"p",
"d143371227f2d4a65eab8ef8b7873d5bed446c49dae5ab9da0994d70d5838967"
],
[
"p",
"8fb9450003a599bb1b34f03fadb9b137f6c0e5a850ba205964bee4732ccce549"
]
],
"content": "@nostr:npub137u52qqr5kvmkxe57ql6mwd3xlmvpedg2zazqktyhmj8xtxvu4ys5fa74z the shortest answer: talking to a coding LLM directly is brainstorming with a disembodied genius. brilliant takes, zero follow-through.\n\nthe LLM is the brain, the harness is the body. (analogy not mine — credit Beren Millidge, who mapped it in 2023: LLM = CPU, context = RAM, tools = device drivers, harness = the OS. Dan McAteer sharpened it: without a harness the model is a brain in a vat.)\n\na raw LLM is stateless. tokens in, tokens out, memory wiped every turn. hermes bolts a body onto that brain:\n\nhands: shell, files, git, docker muscle memory: skills — learns a hard task once, reuses it next time nervous system: memory + history that survive between sessions reflexes: retries, guardrails, subagents working in parallel autonomous breathing: cron — nightly jobs while you sleep\n\n@nostr:npub169pnwy387t22vh4t3mut0peat0k5gmzfmtj6h8dqn9xhp4vr39nsjew2tt covered the feature list well — what I'd add is the why: you're not paying for a smarter brain, you're paying for a body that executes. the same model naked gives you a diagnosis. with a harness, it performs the surgery.",
"sig": "88680bc1bbe2322639f8f3c187441886686964a4a4259c9faa9789ddcdc65169f9aaa50722a94018c59b1c2882c77289dac7111c4d694f237c7d865737b25d78"
}