3 min read13 viewstech

Codex vs DeepSeek Harness vs Hermes: The Real-World Answer Is There's No Best Agent, Only the Right Seat

Romanticist oil painting cover: a lamplit workshop under storm clouds, three blacksmiths each tending a forge — one blazing, one steady, one still smoking — with light breaking through the clouds in the distance

A hands-on video throws three AI agent tools at real work: building a forum theme, fixing bugs, daily ops and fending off an attack. Codex develops best, Hermes runs a month without breaking, and DeepSeek Harness — the prettiest codebase — is the one that crashes. Pairing a local Qwen3.8 27B for grunt work with cloud DeepSeek V4 Flash for finishing cost about 30 RMB for a full day; all-premium would run 100+. Two months of the difference buys a 4090.

  • AI agents
  • Codex
  • DeepSeek
  • Hermes
  • Qwen
  • local AI
  • AI workflow
Contents
  1. What this is about
  2. Three tools, three temperaments
  3. The real savings aren’t in the tool, they’re in the seating
  4. One thing to take with you
  5. Sources

Romanticist oil painting cover: a lamplit workshop under storm clouds, three blacksmiths each tending a forge — one blazing, one steady, one still smoking — with light breaking through the clouds in the distance

A foot may fall short; an inch may excel.
— Qu Yuan, “Bu Ju” (Warring States period)

What this is about

“Which agent is best?” may be the most-asked question right now. A late-August hands-on video answers it without benchmarks — by throwing three tools into the real work of running a forum: building a theme, fixing bugs, daily maintenance, fending off an attack. The answer isn’t a name; it’s a seating chart. Which connects straight to yesterday’s piece: to manage a crew of agents, draw the org chart first.

Three tools, three temperaments

Codex: official, a year-plus old, closed source, the most mature product. In the test it built a forum theme end to end — planned six steps itself, done in twenty-odd minutes; it handles five or six bugs dumped at once, and you can drag screenshots straight in. Its flaw: paired with a local model it over-edits — asked to remove rounded corners, it also flattened the user avatars nobody asked about.

Hermes: Python-based, the least flashy and the steadiest. Left alone without manual updates, it runs three weeks to a month without an error. When the forum came under attack, the host never touched the keyboard — Hermes blocked the attack, restored data, and quietly finished the tagging job every moderator had been avoiding.

DeepSeek Harness: open source, fastest-growing ecosystem (a hundred thousand stars in short order), and the prettiest codebase of the three. It also crashed in the field: one bug it simply couldn’t fix, until the host pushed the code to GitHub and let Codex clean up. The diagnosis is telling — some of its built-in tools it doesn’t know how to call; it can process images, yet insists on being given an image URL.

Three tools, one seat each: Codex in the development seat, Hermes in the ops seat, DeepSeek Harness in the watch seat, each labeled with where it wins and loses

The host’s verdict, plainly: recommend Codex today, use Hermes for daily scripts and maintenance, and DeepSeek Harness — grand prospects, not yet a finished product — is for spare time.

The real savings aren’t in the tool, they’re in the seating

The most stealable move in the video is the cost structure. The token-heavy grunt work — reading code, long development runs — goes to a local Qwen3.8 27B (driven by one AMD 7900 XT, 128K context), free beyond electricity; only what the local model can’t handle goes to cloud DeepSeek V4 Flash to finish. A full day of moderate development plus ops: about 30 RMB. The same load on premium V4 Pro: 100+ a day. His own math: two months of that gap buys a 4090.

He’s honest about the price: the local model is slow (ten minutes on Flash versus half an hour locally) and a notch less sharp (the over-editing incident was the local model’s). So this isn’t “local is cheap, use local for everything” — it’s cheap first, expensive to finish. The same arithmetic as my RunPod piece, except he’s turned the answer into a daily routine.

Cost comparison: top bar mostly free local segment with a small cloud tail, 30 RMB a day; bottom bar all premium cloud, 100+ a day; two months of the gap labeled as one graphics card

One more detail I keep thinking about: the prettiest codebase of the three was the only one that crashed in real work. Elegant architecture and getting things done are different things — judge a tool by whether it finishes the job, not by how it reads. Same as reading company financials: the prettiest slide deck isn’t necessarily the profitable company.

One thing to take with you

The day I watched this I re-seated my own tool list: next to each AI tool I wrote down one of two words — “develop” or “on duty.” Development needs smart, because I’ll see its mistakes; duty needs steady, because when it fails I won’t be there. One of my tools was listed in both columns — that’s the hazard. You could spend five minutes today doing the same: one tool, one seat; the one that fits neither, don’t rush to use.

Sources

Raw Markdown

tech13 views

Mac mini or a GPU for AI at home? First figure out which machine you're missing

A Chinese tech YouTuber explains why Mac minis are selling out: Macs can run language models but struggle with video, and he bought his for stability, low power and a complete software ecosystem, as a desk for directing AI agents. My own setup splits the same way: one machine directs, one old 8GB GPU does the heavy lifting, and training goes to rented cloud GPUs. A plain-language guide to what a control desk needs versus what compute needs. Personal notes, not buying advice.

tech11 views

Picking a GPU for Local AI: The New Cards Worth Buying, and the Old Ones to Skip

Two back-to-back videos from the YouTube channel 掄錘者: one ranks the four cards he keeps recommending, the other talks people out of cheap 48GB veterans. My own Windows desktop runs an older AMD card, and this month I logged what worked and what needed a detour for image generation, video generation, and local language models. A plain-language look at the two specs that matter first: video memory and BF16. Personal notes and test logs, not buying advice.

tech15 views

Does an AI Only Behave When You Swear at It?

A Chinese hardware YouTuber says swearing at AI works: DeepSeek eats it up, OpenAI ignores it. I take his rant apart and find two things inside: pulling the decision back, and asking only for the result. Then I check it against the twenty-three indicator tests I sent out today. No swearing needed, but hand the model a broken ruler and it will politely measure 100%. Education and methodology.

tech12 views

Ten Dollars for One Request: The Day GPT6 Astra Drained Lunchuizhe's Balance, I Went Back and Checked My Own Bill

Lunchuizhe put GPT6 Astra to the test: three chat questions cost a dollar, then one small Codex task turned a seven-dollar balance negative. His verdict was that the model isn't expensive, we're just poor. I pulled a full day of my own AI usage and found the bill tracks how much the model reads, not how much it answers. Educational notes and extension, not a buying recommendation.

tech10 views

A Subscription Beats Pay-Per-Use by a Mile. So Why Won't He Hand It Six Tasks at Once?

Pushed by his commenters, 掄錘者 re-ran the same GPT6 task on a subscription instead of the API: nine dollars yesterday, 4% of a weekly quota today. He conceded the subscription wins, then admitted he no longer dares to fire six tasks at once the way he does with cheap models. I priced my own week of usage at list rates and found that a subscription doesn't save you money. It converts money into a clock that resets. Educational notes and extension, not a purchase recommendation.

investing15 views

Whether It's Worth It Is One Question. Whether You Can Get One Is Another.

A Chinese YouTuber who runs six machines at home reacts to the news that OpenAI bought tens of thousands of Macs to teach AI how to use a mouse. His conclusion: hardware won't get cheaper for two or three years. I checked it against Apple's own earnings call and our memory-supply research. The supply side holds up; the phrase 'this time is different' has a track record in memory. Educational notes and extensions.

Comments

Loading comments…

Sign in with Google before posting. Only your name and profile picture are shown.