A sect reached 390K GitHub stars in half a year — certified by GitHub as "the fastest-growing in platform history." Then, in just four months, its website traffic collapsed from 14.2M visits to 2.77M — a drop of 80%.

That sect is OpenClaw. Early this year, everyone was still "keeping a shrimp." Now, nobody in the jianghu mentions it.

But a crash in traffic doesn't mean it died — it just went quiet, slipping in silently. Count the second-generation sects Chinese companies founded on its techniques, and you'll find more than 30: Tencent's WorkBuddy, Baidu's DuMate, NetEase's LobsterAI, Zhipu's AutoClaw, Volcano Engine's ArkClaw, Alibaba's Qwen Office…

To be precise, its retreat was several things colliding: major bugs shipped in a few big releases (Gateway crashes, memory gaps, security holes — even incidents where local data was accidentally deleted), a brutally high barrier to entry (you had to know the command line, which even spawned a "house-call shrimp installer" business), and its open-source rival Hermes seizing the moment to move users over with a one-click migration command. As for Claude Code and Codex, those great orthodox sects took over the mantle afterward, turning the agent from "something you assemble yourself" into an "out-of-the-box" heirloom.

The traffic didn't vanish — it just changed altars. WorkBuddy hit 20.97M monthly visits, ByteDance's TRAE 12.79M, Alibaba's Qwen Office 7.88M.

Behind all this is a war most people still haven't noticed — the Runtime War. OpenClaw's retreat isn't the end of the story; it's the real beginning. Competition has shifted from "who lets you build your own agent" to "who packages the agent into a sect you can't live without."

This article walks through the war faction by faction — community runtimes, vendor harnesses, platform products — and the endgame hanging over everyone's head.

One spoiler up front, so you don't waste the read: by the end you'll see that the real winner of this war is probably not on today's roster at all. So who is it? Keep reading.

One — First, understand what's actually being fought over

To understand this war, you have to answer a more basic question first: models are already this powerful — why do we still need "the layer in between"?

The answer: a model is an "engine," not a "car." It can reason and generate, but it won't act on its own — it won't open a browser, won't read or write files, won't remember what you told it yesterday, won't pause to ask "is this right?" when it's about to make a mistake. The layer that fills in all these gaps is the runtime layer (Runtime / Harness) — the thing everyone is fighting over right now. Whoever controls this layer controls the valve that turns "model capability" into money.

But here's what you may not have realized: this layer has already been fought over once before — and that time it lost badly. Every single "patch" everyone is frantically adding today traces back to the same corpse. To understand this game, you first have to look down at how that corpse died.

The prequel: how AutoGPT died

In March 2023, AutoGPT blew past 100K stars within weeks and ignited the entire agentic AI movement — agent repos on GitHub grew 900%+ in two years. Everyone thought AGI had arrived. Then, it died.

The causes of death were highly concentrated — and reading them today is especially ironic:

• Infinite loops: the model judged "is the task done?" for itself, fell into recursive verification, and kept searching the same thing over and over;
• Runaway costs: $10–50 per task. One "find a recipe" task burned $14.40 and looped through 47 planning steps only to return nothing;
• Hallucination cascades: one hallucinated step propagated into every downstream step;
• Context amnesia: with a 128K window, the agent forgot what it had done halfway through;
• No reuse: every run started from zero — it had no way to solidify successful actions into a reusable process.

In Amazon's real-world tests, AutoGPT's success rate on shopping tasks was just 24%. And the root cause is a piece of grade-school arithmetic:

This is "multiplicative reliability": every surge of autonomy gets knocked back by this multiplication, then reborn in a more restrained form. And AutoGPT wasn't the first — back in the 1990s, academia already studied "multi-agent systems" (BDI: Belief-Desire-Intention), whose "coordinator + executor" division is the academic ancestor of today's Claude Code Coordinator + Worker; in 2022, the ReAct paper (Princeton + Google) first unified "reasoning" and "acting" into a Thought → Action → Observation loop — every harness's control flow today is a variant of it.

Key judgment: all of today's sandboxes, approvals, HITL, and rollbacks are, in essence, patches on AutoGPT's failure. We discuss "patches" every day but rarely look back at the "wound." Only when you understand the wound do you see why the whole industry is converging in one direction — and what that convergence produces is the three factions below.

Two — Faction 1: "the Resident" community runtimes — OpenClaw, Hermes, Pi

The first faction grew organically out of the open-source community: "personal agent runtimes." What they share: turning AI into a "digital alter ego" that lives beside you — inside your chat apps, ready to do work at any time. The representative is OpenClaw, plus the two others that form a triangle with it: Hermes and Pi. This school is wild by birth — no orthodox temple behind it, just wanderers assembling the frame of their art with their own hands.

OpenClaw: the wandering master who defined "personal agent"

OpenClaw is a local-first personal agent runtime — a persistent Gateway + brain + skills + memory, connected to ten-plus messaging platforms including WhatsApp, Telegram, Feishu, and WeCom. It runs locally, in Docker, or on a cloud server, but by default it's aimed at "just you."

Its biggest contribution is turning "Agent" from a slogan into a concrete thing you can install, configure, and wire to models and tools. It uses file-based long-term memory (SOUL.md, memory/) for near-permanent personalization, and the ClawHub skill marketplace (5,700+ skills) for extension — which is why it's called "the operating system for personal AI." Think of it as a wandering master with no sect — forging the legendary "personal agent," long heard of but never seen, into a real discipline with forms, formulas, and moves anyone can practice.

But OpenClaw has also been doing catch-up work. Its evolution of controllability is the most vivid sample for understanding the whole industry:

Bare autonomy → least privilege (allowlist/deny) → sandbox / approval (HITL) → Policy runtime (a four-way ruling on every action: allow / deny / approve / repair, plus doctor self-checks)

This is a wandering master's full journey from "bare-knuckled brawler" to iron-clad bodyguard: at first he fought shirtless (bare autonomy), then learned to guard himself (least privilege), then donned soft armor (sandbox / approval), and finally forged an impenetrable golden-bell technique (Policy runtime, where every move passes a verdict).

Hidden here is one of the most counterintuitive insights, the approval-fatigue paradox: more approvals means safer, but the more tired the human gets, the more they "rubber-stamp" (blindly approve) — which makes things less safe. So OpenClaw's answer is layered default policies, not "approve everything": low-risk read actions auto-pass, sensitive write actions are gated on approval, and actions that "should never happen" (payments, admin APIs, deleting git history) are denied outright.

And one overlooked prerequisite: auditing is what makes controllability possible at all. "Logging the final tool call isn't enough — you have to log the policy decision itself." The audit log records the full chain of "intent → context → approval → result," not just the final side effect. No matter how strict a jianghu's rules, you still need a ledger recording every move's provenance — otherwise it's a muddled account.

The personal-AI triangle: Pi "understands you," Hermes "gets stronger"

OpenClaw isn't the only species of personal agent. Beside it sit two entirely different beings, and together the three form the triangle of personal AI:

• Hermes Agent (Nous Research, MIT open source): a self-improving agent. Its killer feature is a built-in learning loop — it auto-creates skills from your usage, auto-improves them, with cross-session memory (FTS5 full-text search + Honcho user modeling). In May 2026, its daily token consumption (224B tokens/day) surpassed OpenClaw's for the first time, topping OpenRouter. This is a "self-enlightened prodigy" — where others learn forms from a master, it learns by watching you strike, then devising the next move on its own.
• Pi (Inflection AI): the emotional-companionship layer. It deliberately refuses to execute — no code, no tasks — and leans into EQ: listening, empathy, voice conversation, personal reflection. It monopolizes a dimension no one else can do: emotional continuity. This is a reclusive master who cultivates "inner technique" but no "martial form" — while others compete on who fights better, it competes on who "understands you."

Hermes and OpenClaw are the only pair in the triangle that compete head-on, and every one of their contrasts carries information:

Dimension	OpenClaw	Hermes
Skill acquisition	Catalog model: install manually from ClawHub	Self-learning: observes behavior, writes its own skills
Token economics	Re-scans raw memory — gets more expensive with use	Condenses procedural memory — gets cheaper with use
Architecture	Horizontal: Gateway + multi-agent orchestration	Vertical: single agent, single process
Ecosystem	Deep	Young but fierce

The deepest divide is economic: OpenClaw's value is in "infrastructure reuse" (platform logic), Hermes's is in "compounding time" (organism logic). One is like running a martial academy, earning incense money from disciples; the other is like cultivating alone, growing deeper with each passing year.

And Pi is unique in that it voluntarily exited the capability race to monopolize "heart" instead of "hands" — all three big assistants feel too "tool-like"; Pi is the only one that makes you feel "heard." Watch the 2026 shift: Pi added proactive check-ins, daily briefings, and reminders — companion AI is quietly growing "light execution" too.

There's another phenomenon worth noting: Hermes and Pi are both one-way compatible with OpenClaw — Hermes even ships a hermes claw migrate command that imports an OpenClaw user's settings, memories, skills, and keys in one step. This isn't three parties "embracing each other" — it's latecomers actively matching the first mover to inherit its users. What actually makes this compatibility work is something neutral underneath: the agentskills.io skill standard initiated by Anthropic (alongside MCP and Markdown memory). Skills are therefore becoming portable assets across runtimes, like containers were to cloud computing. The "complete personal AI" at the center of the triangle probably won't be built by any single player, but stitched together by this open-standard layer. It's like the jianghu where each sect's kung fu differs, but the fundamental methods of "inner power," "forms," and "mental technique" are being unified under one shared "manual of martial arts."

But just as these "folk experts" thought they were defining the future, they turned around and found — the great orthodox sects had already descended the mountain, wielding guns they themselves invented. The next chapter is where everyone's spine goes cold.

Three — Faction 2: "the First-Party" vendor harnesses — Codex, Claude Code, DeepSeek, and the Chinese giants

If community runtimes are the first faction, the second is sharper — model makers building their own harnesses. Overseas it's Codex and Claude Code; in China, it's the collective all-in of four Chinese giants. While the wanderers were still training in the wilderness, the Shaolin and Wudang of this world had already come down the mountain with blades drawn.

Codex and Claude Code: coding is only the beachhead

Codex and Claude Code are often treated as "programming tools," but coding is only the beachhead — because code has verifiable feedback (compile, test, diff), it's the easiest place for an agent loop to prove itself. They're rapidly generalizing. Dify's official blog even admitted its new agent system copies Codex and Claude Code's pattern (ReAct loop + Bash + sandbox).

The sharpest weapon in their hands is vertical integration. Codex's engineering lead, Thibault Sottiaux, has a quote the industry quotes relentlessly:

"Scaffolding is coping, not scaling."

Unpacked, it's a three-hit combo:

1. Model + harness co-training — using the No Free Lunch theorem (optimizing for a specific distribution beats generalizing across all distributions), harness problems can be fixed directly in the next model generation;
2. Three layers in one body — the base model, the agent framework, and the user product all live in one organization; middle-layer players are all stacking harnesses on "someone else's model";
3. A two-way flywheel — real usage data flows back into training, extending scaling laws from the model layer to the system layer.

In jianghu terms, these three moves mean: an orthodox sect's inner power, forms, and disciples all live under one roof — trained inside and out, and endlessly replenished by real combat.

And talent harvesting is already a fact: in February 2026, OpenClaw founder Peter Steinberger joined OpenAI's Codex team.

The most talented creators in the personal-agent space are being absorbed one by one by the model companies. This is the deepest worry of the community runtimes — the paradigm was theirs, but the talent and the standards are flowing toward the vendors. The wanderer founded a sect, and in the end even its own sect-leader was invited away as a guest master of the orthodox school.

The difference between the two routes is also clear:

	Claude Code	Codex
Philosophy	Plan-heavy ("think first"), winning on plugin & community skill ecosystem	Precise execution + skill packaging, distributed via the OpenAI stack
Key moves	Projects rebuilt as a Coordinator + parallel worker multi-agent architecture, with shared memory	Record & Replay: demo an operation once on screen, auto-generate a reusable SKILL.md
Ecosystem weapon	MCP (the protocol Anthropic pioneered) + deep plugin & skill ecosystem	ChatGPT distribution + an Apache-2.0 open-source CLI
Hidden worry	Sandbox off by default; permission & prompt-injection risk in autonomous runs	MCP ecosystem less mature than Claude Code, and locked to the OpenAI model family

And Claude Code has a "shadow" you can't ignore — OpenCode. It's often called "the open-source second choice to Claude Code": also a terminal-native agent (long past just coding — a general agent that tackles all kinds of tasks), but one that consistently ranks #2 behind Claude Code in developer polls (200K+ stars, 16M developers a month). The difference is that Claude Code is Anthropic's official closed harness locked to Claude, while OpenCode is the open-source, auditable one that plugs into 75+ model providers — the choice for engineers who don't want to be locked to a single vendor. This pairing is a microcosm of the whole runtime war: vendors use closed harnesses to bind their own models; the community answers with an open harness as its replaceable mirror. This is exactly the jianghu's contest between "orthodox disciples" and "wanderers who join with their own skills" — the orthodox sect takes only inner disciples sworn to its own kung fu, while the wanderer takes any swordsman willing to learn from whoever has the skill.

The Chinese giants' four-way war: Tencent, ByteDance, Alibaba, Baidu

On the Chinese side, it's the collective all-in of four Chinese tech giants. Huxiu's summary is sharp: "Tencent focuses, ByteDance spreads, Alibaba balances, Baidu races." They're actually fighting on two tracks:

• C-end general assistant track (whoever has the most users): Doubao at 419M MAU is the runaway leader (QuestMobile 2026 Q3, including ecosystem-matrix traffic), followed by Qwen at 158M, DeepSeek at 123M, and Yuanbao at 88.7M — this round, ByteDance won on volume through Douyin's distribution;
• Office-agent track (whoever actually gets the work done): Tencent's WorkBuddy leads at 32.35M MAU, Baidu's Dazi is second, Qwen Office third; ByteDance's Doubao Work only entered at the end of August, the latest.

The four giants' strategies map to four different entry logics:

Player	Office agent	Entry / ecosystem	Strategy
Tencent WorkBuddy	32.35M MAU (top)	WeChat / WeCom	Focus: bet on one, shut down QClaw
ByteDance Doubao Work	Only launched end of Aug	Douyin / Feishu	Spread: Doubao Work + TRAE dual track
Alibaba Qwen Office	Third (millions)	DingTalk / Taobao / Alipay	Balance: the only one in top 3 on both C-end and office
Baidu Dazi	Second (fastest growth)	Search / Xiaodu hardware	Race: 150 iterations in 5 months, but lacks enterprise-collab entry

These four are the jianghu's four great lords — Tencent the prudent, steady orthodox official sect (WeChat is its dragon vein), ByteDance the rich and populous magnate (Douyin is its million-strong army), Alibaba the seasoned, balanced grandmaster (DingTalk, Taobao, and Alipay as its three heads and six arms), and Baidu the former overlord racing to reclaim its crown (search is its old trade — and its shackle).

Among them, WorkBuddy has the most direct connection to our OpenClaw storyline, and deserves a word on its origin. Tencent's official position: WorkBuddy is a self-developed desktop agent, from the same lineage as CodeBuddy; but as multiple media outlets have confirmed, it was indeed "inspired by the OpenClaw open-source framework" — a Tencent team of about 10 people built the initial version in two days. Its play is the third road of "enterprise-izing OpenClaw": compatible with OpenClaw's Skills ecosystem (SkillHub lists 70,000+ skills) to harvest ecosystem dividends, self-built security gateway/permissions/audit to keep control, and WeChat/WeCom/QQ channels for distribution. A master of the orthodox official sect spotted the wanderer's wild-road kung fu and, in just two days, "tamed and refined" it into the court's standardized saber style — borrowing the wanderer's momentum while keeping its own control.

What's truly worth watching about WorkBuddy among the four is two data points most people overlook:

First, its user-stickiness lead is a chasm. Xsignal data shows WorkBuddy users average 20.2 minutes and 19.5 sessions per month; Baidu's Dazi only 2.5 minutes and 4.7 sessions, Qwen Office 9.1 minutes and 5.2 sessions. Others are "open it, ask one question, leave" — WorkBuddy is "sit down, open the document, actually do the work." That's also why it dared to be the first to charge (¥99–999/month) in July: it's not selling "chat," it's selling "deliverables."

Second, it hits the hardest need of government and big enterprises. WorkBuddy emphasizes local file processing, structured deliverables, and controlled permissions — exactly the first hurdle for AI adoption in data-security-sensitive scenarios (government, finance, large enterprises). Huxiu nails it: the competition metric for office agents is shifting from "MAU" to "task completion rate and delivery quality," and data security is the main barrier to enterprise adoption.

But to be honest about its short board: it can't beat ByteDance and Alibaba on C-end volume. Doubao's 419M MAU is 13× WorkBuddy's, because it rides Douyin — a national-scale distribution machine — while WorkBuddy's battlefield is "boxed in" to office, a narrower but more valuable scenario.

So the conclusion: WorkBuddy's special advantage isn't "more users," it's "users who stay + enterprises that dare to adopt + a WeChat entry that's already open." It's fundamentally a staggered competition — ByteDance won the "national entry," Tencent won the "office entry"; the former is traffic logic, the latter is task logic. And the point we've repeated throughout — that the deciding factor of the runtime war is governance — gets a Chinese case study right here in WorkBuddy.

DeepSeek / dsh: the third road among the vendors — turning the harness into an operating system

Among model vendors building their own harnesses, DeepSeek is the most special one — it's both the "foundation" and the "arms dealer," and a challenger that doesn't follow the usual path. Remember the "sweeping monk with no sect" from our Hundred Models essay? This is him. DeepSeek was the one lone figure in the great-model melee who joined no sect yet carried the deepest inner power — and now it's applying the same playbook to the harness layer.

First, the model layer: DeepSeek V4 (Preview released April 24, 2026) — million-token context, MIT open weights, unified reasoning and chat, SWE-bench Verified 80.6%, #1 open-source. Its most famous feat is undercutting high-performance Chinese models on price — V4-Pro's total cost is about 1/7th of GPT-5.5. In one line: it's the default engine for nearly every middle-layer player. The sweeping monk never steps into the ring — but his inner power is silently borrowed by every sect-leader in the jianghu.

But it has a fatal paradox: it supplies ammunition to the whole industry, yet controls no runtime-layer entry point — its value gets skimmed by the harness-layer players.

On August 13, 2026, it gave its answer: open-sourcing dsh (DeepSeek Harness) — 10K stars in half an hour, 50K in 12 hours, 170K in a week. Its philosophy is pointed straight at Codex:

	Codex / Claude Code	dsh
Core thesis	Complex scaffolding is a cover-up; co-train model + harness	No privileged kernel; even the agent loop is a plugin
Form	A tuned, finished car	The strongest skeleton + all the parts — you assemble
Replaceability	Low (vendor hardcodes the kernel)	Extreme: sandbox/approval/filesystem/model/memory are all "seams," each replaceable
Auditability	A product feature	Runtime law: "what the model sees is recorded"; violation triggers an alarm

The orthodox sects teach "complete, finished arts"; the sweeping monk posts the original manuscript of the kung-fu manual directly on the mountain gate — every line of the inner-power formula is yours to take apart, rewrite, and start a new sect from.

The deepest layer is strategic: DeepSeek's vertical integration happens in the "data layer," not the "product layer." The harness's core value is "turning model capability into stable end-to-end delivery while accumulating the scarcest data — long-horizon task trajectories and error-correction feedback." OpenAI locks in users to buy data; DeepSeek uses open architecture to buy it — dsh doesn't lock users, but every session's log feeds the next model's training data. Other sects keep people by taking inner disciples; the sweeping monk locks no one — but every trace of your training becomes the raw material for his next layer of insight.

But the cold start has three hard thresholds: product maturity (currently a v0.1 preview — a skeleton, not a finished car), reliability math (multi-step tasks multiply errors; discovering, containing, and rolling back failures takes time to polish), and compliance tickets (from 2027, banking procurement catalogs require MLPS Level 3; Zhipu ZCode, Tencent CodeBuddy, and Alibaba Qoder have already staked claims, while dsh has no certificate yet).

Judgment: dsh's endgame is "the Linux of the agent layer" — if it matures, it will "infrastructuralize" the execution layer, forcing closed-source vendors upward to sell end-to-end workflow integration value. DeepSeek's real moat is in the data layer: a fully open stack (self-hosted V4 + dsh) lets European enterprises get 1/10th the price with full sovereignty. But success hinges on execution, not philosophy — it can afford to lose the architecture; it can't afford to lose the time.

But there's still a batch of players that never planned to "stay by your side" — what they want is to turn the agent directly into a finished product you can deliver and sign a contract on. On this road hides the most expensive company in the game, and the most brutal takeover battle.

Four — Faction 3: "the Finished Product" platform products — Dify, Coze, Manus, Devin

The third faction takes an entirely different road — turning agent capability into a "deliverable, governable product platform" rather than a runtime that lives beside you. On this road, four players each represent a form: Dify is the "workflow platform," Coze the "low-code platform," Manus the "general finished agent," and Devin the "coding finished agent." The first two factions are still fighting over "who stays by your side"; this faction simply opened an escort agency instead — no companionship, just take the order, deliver the goods, cash on delivery.

Dify: guardian of the workflow, and OpenClaw's mirror image

Dify is an LLM application development platform (LLMOps): it lets teams visually orchestrate workflows on the server side, stringing models, RAG, and tools into deliverable AI apps that go out via API. Its core difference in one sentence: Dify is "humans draw the flowchart, the model fills in the blanks"; OpenClaw is "the model decides its own next step inside a loop, with tools as its hands and feet."

The two stand on the same middle layer; the difference is where control flow lives:

Dimension	Dify	OpenClaw
Control-flow owner	Human pre-arranges a DAG; LLM is a node	LLM plans in real time inside the agent loop
Deployment	Server / cloud SaaS / on-prem, team collaboration	Local / Docker / cloud server, defaults to personal
Memory model	Session state + RAG knowledge base	File-based long-term memory (AGENTS.md, sqlite-vec)
Interaction entry	API / embedded apps (for programs)	Messaging platforms (for people, always-on)
Design philosophy	Determinism & governability first	Autonomy & capability ceiling first

One is an orthodox sect that lives by "rules" — every form is scripted, every form is accountable, every move is recorded; the other is a wanderer who lives by "instinct" — improvising on the fly, striking however it pleases.

But the most striking thing isn't how different they are — it's that these two arch-enemies are moving toward each other in mirror symmetry. OpenClaw pulls back from "bare autonomy," growing controllability (Policy, approval, audit); Dify pushes outward from "rigidity," growing autonomy — its four levels of self-evolution:

Pure deterministic workflow (DAG) → HITL (human-in-the-loop) → sandbox Agent + Skills → Build Mode (conversationally build an agent; changes land in a Build Draft)

Dify's most ingenious design is turning "self-evolution" into "controlled evolution": in Build Mode, the agent works and modifies as it goes, but changes don't take effect directly — they land in a Build Draft + build_note.md, and you adopt or discard each one in the config panel, publishing only when satisfied. Dify allows evolution, but evolution must be reviewable, rollback-able, and human-approved. An orthodox sect allows its disciples to innovate too — but every new move must first be written into the "manual of forms," reviewed by the sect-leader and critiqued by fellow disciples, before it can be taught outward.

It also draws a crisp line between agents and workflows; the official positioning is remarkably clear-headed:

The agent decides "how to complete the task"; the workflow decides "when the task happens, what comes before and after, when a human must step in, and what to do when something goes wrong."

Dify repeatedly defines an agent's value with one phrase — "your own software patch": agents are good at improvisation, but "one success isn't a system"; workflows turn improvisational ability into a repeatable, schedulable, monitorable, stable process. And one line from Dify's official team is almost a confession: "Earlier this year when OpenClaw went viral, we could have rushed out a Dify version of it. We didn't." — its agent rework was pushed along precisely by the CLI-agent paradigm that OpenClaw represents.

Put the two side by side and the conclusion is clear: autonomy and determinism are not a zero-sum seesaw, but orthogonal capabilities that can stack. Both sides have walked to the middle ground of "sandbox + approval + auditability"; the only difference is the default order of gas and brake.

Coze: the warning of a de-emphasized workflow

A lot of people panicked when they saw Coze de-emphasizing workflows: are workflows about to be replaced by agents? The answer will surprise you — it's the exact opposite.

First, admit the fact: Coze is indeed de-emphasizing workflows. And it's not a feature tweak — it's four major pivots of the entire product positioning: Bot Studio → workflows → Coze Space → Coze Coding / folded into Doubao. The homepage hero copy has already become "tell Coze what you want to accomplish, and the agent will take over with materials and tools." The workflow wasn't deleted, but it retreated behind the scenes, generated by conversation. This is the jianghu's "treasure-pavilion shopkeeper" — backed by ByteDance's mountain of gold, he shelves every treasure and dares to stock every novelty, all to spread the stall wide and keep every passing customer; what he lacks isn't capital, but a single make-or-break signature art of his own.

But the conclusion is exactly the reverse — the workflow won't die; what dies is the act of "hand-dragging nodes." The key difference:

• Coze's workflow is a "feature" — an editing play for C-end user experience. When agents get strong enough that users don't need to draw, this experience loses its value.
• Dify's workflow is an "asset" — the vehicle of accountability on the enterprise side. Its value was never "fun to drag," but the artifact itself: who changed it, why, can it roll back on error, can someone be held responsible in an accident.

The best analogy: did AI writing code kill the IDE? No. What got replaced was the act of "hand-writing code"; what became more important was Git, Code Review, CI/CD. Likewise, after agents generate workflows, what gets replaced is "hand-dragging nodes," and what matters more becomes diff, approval, versioning, monitoring. The treasure-pavilion shopkeeper trains to "fill the shelves and gather the crowd"; the orthodox sect trains to "found a school and pass it down a hundred years" — goods aren't worth money; accountability is.

Because the artifact is the accountability unit — no matter who wrote it.

The workflow is the "source code" of enterprise AI applications. AI can write source code, but source code itself won't disappear — it just changes from human-written to AI-written, human-signed. So: what Coze loses is "the workflow as a plaything"; what Dify defends is "the workflow as a contract."

But Dify isn't without risk either. Its real risk isn't "the workflow being replaced," it's who absorbs the governance layer: being absorbed by model vendors, by agent harnesses, or the performance debt (workflow-engine overhead is a real bottleneck under high concurrency). Its hedging moves are clear — Build Mode (joining the generation wave) + sandbox Agent (joining the autonomy wave) + transforming the workflow into a coordination layer.

Manus: the general "finished-delivery" agent, a capital specimen that changed hands three times

If OpenClaw gives you "a runtime living beside you," Manus walks the exact opposite road — a cloud employee: you toss it a goal, it runs asynchronously, and hands back the finished product along with its entire process. It doesn't "work with you"; it "finishes the job for you." This is the jianghu's "wandering swordsman who can't steer his own fate" — full of skill, yet he changed hands three times, and couldn't even decide his own home.

In March 2025, Manus blew up overnight: a demo video of "autonomously screening resumes, analyzing stocks, generating reports" pushed invitation codes to ¥100K each on the secondary market, and its site topped 10M visits within 4 hours of launch. It hit SOTA on the GAIA benchmark and claimed to be "the world's first general AI agent."

But what truly made it a specimen is its capital fate — 16 months of an "acquired — blocked — bought back" rollercoaster, with its valuation doubling:

• December 2025: Meta closed the deal in just 10 days, acquiring it for about $2B (one of Meta's biggest AI acquisitions at the time); founder Xiao Hong became a Meta VP;
• April 2026: Chinese regulators blocked a foreign acquisition in AI for the first time, ordering the deal revoked;
• July 2026: Tencent led Chinese capital to buy it back at the same $2B valuation; Manus now operates independently in Singapore, eyeing a Hong Kong IPO;
• October 2026: closed over $500M in new funding (led by Boyu and IDG), targeting a doubled $4B valuation.

Backing this capital frenzy is hard revenue: Manus crossed $100M ARR within 8 months of launch (the fastest on record globally), and by July 2026 had hit $400–500M. It proved the "general finished agent" form has real demand — but it also exposed a brutal reality: even a product this strong and this well-funded can have its whole story rewritten by a single regulatory veto. Beyond technology, market, and capital, there's a higher hand moving the pieces. No matter how fast the wandering swordsman, he can't outrun a single court decree.

Devin and Cursor: two philosophies of the coding agent

In faction 3 there are also two players often lumped together with Codex and Claude Code, but who actually belong to "another road" — they turned the act of "writing code" itself into a deliverable product. These two are the jianghu's unorthodox swordsmen — while others fight over "whose sect is biggest," they draw their blades in the bloodiest, most dangerous arena of all: code.

Devin (Cognition) is the "world's first fully autonomous AI software engineer": it takes a complete task, writes the code, debugs it, and delivers it — all by itself. It walks the autonomous-agent route. Its hard numbers are stunning: 1M+ MAU, ARR climbing from $492M to $900M in just four months, and a $2B Series E in September 2026 at a $48B valuation. The most telling detail — Cognition claims Devin already writes 90% of its own code internally: a company making AI programmers, using its own AI programmer to build its product. This is "sharpening your sword with the very sword you forged."

Cursor walks another route — collaborative editing: it embeds AI in the editor for completion, refactoring, and explanation, with the human always in the loop. It reached $2B in annualized revenue in three years, and was acquired by Musk for $60B into SpaceX (August 2026).

Put these two side by side, and it's the "autonomy vs collaboration" spectrum we've been tracing, replayed in the coding domain:

	Devin (autonomous agent)	Cursor (collaborative editing)
Control	Hand over the whole task; human only checks results	Human stays in the loop; AI does completion & refactoring
Philosophy	Autonomy first (like OpenClaw)	Determinism first (like Dify)
Valuation / price	$48B valuation (Cognition)	$60B acquired by SpaceX
Shared worry	A general coding agent can't replace the Code Review step of "understanding the code, judging right from wrong, and taking responsibility for launch" — AI can write code, but "accountability" still falls to humans

Coding is "the battlefield most willing to count money" in this war for a reason: code can directly compute ROI — save one engineer, and that's hundreds of thousands in annual cost, which finance can write off directly. So while other agents are still figuring out how to get users to pay, Cursor has already hit $2B ARR and Devin a $48B valuation. And that's why coding became the agent's first beachhead — not because "code is simple," but precisely because "code is the most valuable and most verifiable." In other words, the swordsmen drew their blades there first not because it was easy to fight, but because the bounty there was highest.

Three factions, done. Now comes the harshest question — if one day models absorb all these "middlemen" by themselves, will the layer everyone's fighting tooth and nail over turn into a bubble?

Five — The endgame: the dissolution thesis is everyone's shared exam

Finally, pull the view out to a question no one can get around:

If models get strong enough to internalize tool-use, planning, and memory, will this entire layer slowly degrade into a hollow shell — like the IDE of old?

This "dissolution thesis" is the question everyone is dodging yet no one can escape — it simultaneously threatens and props up every player. In jianghu terms: if inner power could one day generate its own moves, would all the academies and sword-schools dotting the mountains suddenly find themselves out of work overnight?

The pro side's evidence is hard: the harness is indeed "thinning."

Year	Form	Reason
2023	Heavy frameworks (LangChain/AutoGPT/CrewAI)	Models unreliable, need rails to guide them
2025	Harness (Claude Code/Codex)	Models smart enough; just add sandbox + tools
2026	Thinner harness (Browser Harness, "no framework")	Stable tool calls, million-token context, model self-healing

Browser Use co-founder Gregor Zunic's line in April 2026 is the footnote to this trend: "We got tired of frameworks constraining LLMs, so we deleted the framework." He also threw out a harsher claim: "An agent is nothing more than a for-loop over messages." With every step, there's one fewer abstraction layer between the model and the tool.

But the con side's evidence is harder. Terraform founder Mitchell Hashimoto dropped a formula in 2026 that became industry consensus:

Agent = Model + Harness. The model is just the engine; the harness is the control plane that turns the engine into a "governable vehicle."

AWS went so far as to make "Managed Harness" an official feature of Bedrock AgentCore in April 2026. The con side has three hard arguments:

1. The stronger the model, the more important the harness — because you can hand it more dangerous, longer-horizon, fuzzier tasks, and those tasks demand stronger policy, sandbox, approval, and observability;
2. The word "agent" has stopped meaning anything — the Harness formula gives the industry a boundary it can design against, sell against, and hire against;
3. Production data slaps back — according to multiple 2026 industry reports, only 5–12% of enterprise agent projects actually reach production; most die in the demo-to-pilot gap (missing security review, observability, and governance models).

So the dissolution thesis's real conclusion is:

What sinks into the model is "intelligence" — reasoning, planning. What rises into infrastructure is "governance" — policy, sandbox, audit, observability. What engineers were doing in 2026 is dismantling the "smart logic" they wrote a year ago (because every new model makes it obsolete), while keeping five layers that won't dissolve:

• Guides: constrain what it knows and what it's asked to do;
• Policy / Gateway: constrain what it's allowed to do, no matter what it thinks;
• Sensors: monitor drift, loops, bad output;
• Orchestration: session isolation, retries, micro-VMs;
• Observability: record every decision, debuggable and auditable.

Notice there's no "model" in that list — deliberately: the platform layer and model layer are designed to be independently replaceable. This is the truth of the jianghu: martial arts may be lost, masters may fade, but the "rules of the jianghu" — the sects, the rosters, the precepts, the referees of duels, the ledgers of grievances — have been passed down generation after generation and never once vanished.

Why is this "everyone's shared existential question"? Because it threatens and props up at the same time:

• For Dify: the "smartness" of workflow orchestration gets eaten by the model; what remains is governance and assets — exactly what it's betting on. The dissolution thesis is both a threat and a validation;
• For OpenClaw/Hermes: the agent loop gets internalized into the model, shrinking runtime value — but memory, channels, and ecosystem are things the model can't give;
• For Codex/Claude Code: vendor harnesses are closest to the "dissolution" definition, yet even they keep sandboxes, Policy, and Security — even the most aggressive vertical integrators haven't dared to believe "model = agent";
• For dsh: it's simply betting on "harness persists, but every layer is replaceable" — not on whether the model eats the harness, but on the fact that whoever wins still needs a neutral, replaceable, auditable execution base.

The truth of the dissolution thesis isn't "the harness will disappear," but "the harness's intelligence will sink into the model, and the harness's governance will rise into infrastructure." Just as compilers weren't replaced by languages, and operating systems weren't replaced by applications — they merely handed over the "smarts that go obsolete" while gripping the "control that never does" even tighter. A grandmaster's ultimate technique may be lost to time, but the name "Shaolin" or "Wudang" will always have someone guarding it.

In closing

Fold this whole board up and it's seven sentences:

1. There is no multiple-choice "who represents the future." The shape of the future is three layers: model layer → runtime layer (execution + governance) → user. All the players above are different slices of that middle layer, not two rival candidates.
2. Autonomy and determinism aren't zero-sum; they stack. The whole industry is converging on an "autonomous + controlled" hybrid — only the default weights differ.
3. "Self-evolution" is the watershed. OpenClaw/Hermes's evolution is runtime-online (the model changes behavior in real time); Dify's is offline-controlled (changes go to draft, humans sign off, rollback-able).
4. They won't merge. Because "who holds the steering wheel" is two irreconcilable product philosophies. In self-driving terms: L2 and L5 both converge technically, but the watershed of "does a human still hold the wheel" never disappears.
5. Skills are becoming the most universal asset unit. From Codex's video recording, Hermes's auto-observation, OpenClaw's marketplace, Dify's manual packaging, to Tencent's SkillHub — the only difference is how they're acquired.
6. The prequel reveals this round's essence. The death of AutoGPT and the birth of ReAct show this isn't the first attempt at autonomous agents — it's the first time "multiplicative reliability" has a chance of being solved head-on with engineering.
7. The dissolution thesis reveals the endgame's shape. The runtime layer won't disappear, but "intelligence" will sink into the model and "governance" will rise into infrastructure. Whoever makes governance most trustworthy holds the position in the next war.

This explains why Dify, OpenClaw, and dsh all, without coordination, upgraded "auditability" from a feature into a piece of "runtime law."

Because everyone has seen the same thing:

AI can write the contract, but the contract itself won't disappear — it just changes from human-written to AI-written, human-signed.