What Are We Talking About When We Talk About Harness? | In-Depth Conversation: MiniMax × Hermes Agent
Summary
OpenClaw broke through the adoption barrier: Chinese users had previously lacked a usable agent; models such as MiniMax M2.5 and M2.7 filled the capability gap, while IM access, Spring Festival traffic and ByteDance’s Spring Festival Gala investment in Doubao combined to produce a near “zero-to-one” leap in user experience. The boom is not simply fading: after the initial education phase, users are now demanding stability, memory and sustained delivery, shifting attention from “raising shrimp” to “taming horses.”
Tommy puts it plainly: “The model is the brain; the agent framework is the hands.” From installation to first task takes less than 2 minutes; successful workflows can also be distilled into skills, so an agent does not get a task right once and fail the next time. The same framework and skills can deliver the expected output across 8 models; Tommy recalls daily token consumption on OpenRouter rising from 2B to 20B in just over a month, with yesterday’s figure nearly reaching 300B.
Harness is evolving from a product feature into a production system: it gives models tools, environments and latitude while using permissions, feedback and peer review to rein in the wild horse. Its value is not role-play, but letting agents decide whether they can deploy, test, launch an A/B test, and turn feedback into skills, CUI or hooks. Adao’s judgment is that “humans have become the bottleneck”; the key to future work is not running more agents, but closing the loop on steps that previously required human confirmation.
Multi-Agent’s gains come from fresh context and cross-validation: people often reply “approved, keep going” or “you got this,” while 2 models can exchange complete arguments at high density. More importantly, once a single agent’s context exceeds roughly 50%, intelligence may decline exponentially; after veering off course, it also tends to drift further. An independent agent provides a second budget and a completely new context—a cold shower that improves the reliability of long-horizon tasks.
The moat around the general-purpose application layer is thinning. Model companies will absorb the techniques users unlock, and workflows, skills and successful execution traces may ultimately become training data; Adao says general-purpose agent applications “will always be internalized by the models.” More defensible positions lie in the industry last mile, compliance responsibility, proprietary know-how, and tools and environments built specifically for agents. Identity and payments infrastructure, meanwhile, are more likely to accrue to major platforms than to ordinary startups.
The China-US gap is not just about training recipes. Adao believes the gap in training methods and model understanding is no longer that large; the real difference lies in defining real-world tasks, organizing top industry experts to “teach the model,” co-building harnesses with enterprises, and securing the compute needed to scale. Opus 4.7 is viewed as a stronger RL model, but optimizing only for final outcomes may produce process hallucinations and weaker constraint adherence, showing that better benchmarks do not automatically translate into controllable delivery.
The end state looks more like the interface disappearing than people disappearing: Manus-style result delivery, local long-running agents and bring-your-own-token models currently serve different users, but Adao expects them to converge into a multimodal, near-real-time assistant that requires only necessary interaction once models are strong enough. “The outer layer is getting thinner and thinner.” Humans move up the stack toward goals, taste and creativity; unit intelligence costs continue to fall exponentially. Not using AI will be like bringing a melee weapon to a gunfight—Adao’s practical conclusion is: “You can’t beat them; you have to join them.”
Deep dive
1. OpenClaw Gave Chinese Users Their First Agent Leap
Adao recalls that OpenClaw began gaining traction overseas around January. On the day MiniMax went public, the team was dealing with Anthropic cutting off its subscription access and contacted Peter to find a suitable model for the framework’s extreme token consumption. The project later helped drive Mate Mini to a sellout, but the team initially did not expect it to become so popular.
Domestic enthusiasm even exceeded Silicon Valley’s, and not because of marketing alone. Overseas users had already encountered Claude Code and Cowork, while Chinese users lacked an agent at a comparable level. Once models such as M2.5 and M2.7 acquired agent capabilities, and IM lowered the access barrier, “that last sheet of paper” was finally torn through.
Koji referred to the “everyone raising shrimp” craze. Adao added that ByteDance’s all-out Spring Festival Gala investment in Doubao, combined with OpenClaw, moved mass-market AI awareness from chat tools into the Agent era.
2. Hermes Used Memory to Meet the Reliability Demand After “Raising Shrimp”
Zeying observed that the cycle from OpenClaw’s breakout to normalization and then Hermes’ rise took roughly 1 month. Users had experienced AI productivity gains, but also instability and the frustration of “refreshing memory at 4 a.m. every day”: an agent painstakingly tuned the day before might ask the next day, “What did we talk about yesterday?”
Tommy’s definition is direct: “If the large language model is the brain, the agent framework is the hands.” Hermes handles tool orchestration, the main loop, state and error handling, allowing the model not merely to answer questions but to execute in the real world and deliver outcomes.
A 2-minute cold start is the product’s first advantage. Through the Nodes platform API, users can go from installation to their first task in under 2 minutes without wrestling with bloated code or complex configuration, and can deploy it on their own computer or a VPS.
More important is the memory-skill loop: successful workflows are saved as skills and can then be reproduced reliably. Even across 8 different models, the same framework and skills can maintain the expected output. Tommy’s “self-evolution” is fundamentally not a personality myth, but knowledge compression, cross-model consistency and trust.
3. Nous Grew from an Open-Source Community into a Pragmatic Product
Tommy explained that Nous Research began as a Discord community in 2022. A group of people who enjoyed open source and tinkering with models gathered around Llama post-training; the first Hermes models tried to move away from the canned assistant voice, making text more human and more varied.
The research later expanded into YaRN, extending context from 4,000 to 12,800. The team also explored distributed training across non-co-located GPUs through DisTral, aggregating scattered compute while continuing to train open-source models.
Hermes Agent was born out of an “extreme focus on pragmatism.” About a year ago, co-founder and head of post-training Tegniam simply wanted to manage his daily work; benchmarks were not the starting point. Demand after the open-source release far exceeded expectations. Tommy recalls daily token consumption on OpenRouter rising from 2B to 20B, with yesterday’s reading nearly reaching 300B.
4. Harness Gives a Wild Horse Freedom—and a Bridle
Zeying said harness technology can be divided into 6 layers, but is more intuitive as the equivalent of hiring a colleague: define permissions first, then provide tools such as a model, notebook, phone and email. The core principle is to constrain and authorize at the same time, allowing the agent to deliver complete outcomes within clear boundaries.
A single agent is still not trustworthy enough, so multiple agents can be assigned to produce, review and challenge work separately. One might generate the output while another looks for problems and audits the work. The goal is not to imitate job titles, but to use adversarial objectives to improve results when no human is in the loop.
Adao offered 2 metaphors. The model is the powerful engine of a Gundam, while the harness is the mech that lets the engine reach its potential. Or the model is a highly capable wild horse: “The question is how to construct the bridle that lets it perform at its best.” Once model capabilities, tool permissions and real execution conditions mature, the harness emerges naturally.
5. The Bottleneck Has Moved from the Model to the Feedback Loop
Adao said that by last September and October, his workflow had almost no use for an IDE: 5-6 agents ran concurrently on his local machine, while another 10 agents worked in the cloud across Sandbox and GitHub branches on different approaches. The problem instead became that “humans have become the bottleneck,” because he had to keep switching contexts and adding instructions.
The solution is to give agents real feedback: let them judge whether a program can be deployed to production, whether the result is correct, run tests in practice and even launch an A/B test, then revise the path based on the feedback.
These experiences can be distilled into skills, CUI or hooks, gradually automating checkpoints that previously required human confirmation. Harness is therefore not a UI layer wrapped around the model; it connects planning, action, verification and learning into a sustainable loop.
6. Multi-Agent Uses a Second Context to Fight Long-Horizon Drift
In response to the question of why a strong model still needs an agent team, Zeying pointed to information density. A model may produce a long “essay,” to which a human responds only “approved, keep going” or “you got this”; 2 models can exchange large amounts of information efficiently, giving collaboration higher throughput.
The harder limit comes from context. Zeying said that once a single agent’s context text exceeds roughly 50%, intelligence may decline exponentially; even the strongest model cannot keep talking indefinitely. Multi-Agent breaks the task into new contexts instead of letting one conversation drag down the entire reasoning process.
Adao cited the conclusion of a long-horizon task study: when the route is correct, an agent keeps getting things right, but once it deviates, it tends to drift further off course, like a person fixating on an idea. A second agent can cross-check from another angle in an entirely new context.
Koji compared it to someone suddenly throwing a bucket of cold water over you. Adao’s version is, “Go sleep, then use a different brain.” The value is not the 2 budgets themselves, but that the second budget has an independent starting point and responsibility for verification.
7. Self-Evolution Starts on the Model Production Line
Adao believes “self-evolution” attracts more attention than recommendation algorithms because large-model training is expensive and organizational collaboration breaks down rapidly as headcount rises. To scale, companies must reduce hands-on human operations and have Agents and AI take on more of the work.
Later-generation models depend on earlier ones: older models clean and search pretraining data, construct datasets and build RL environments. Adao estimates that roughly 70%-80% or more of the work in M2.7’s RL pipeline is already handled by models and agents.
The remaining human work is no longer checking experimental failures one by one. It is reviewing aggregate results, comparing them with the intended direction, discussing agent recommendations and ultimately deciding where to go through judgment, taste and creativity. “Humans are the ones steering,” not exiting the system.
8. Reliability and Long-Term Operation Change the Human-Agent Relationship
Koji asked whether agents running for longer periods provide the foundation for self-evolution. Adao corrected the framing: more important than duration is being reliable. Continuous learning only matters if the system can consistently solve problems that are sufficiently complex and long-horizon.
Zeying explained ordinary users’ needs through cold start. At the first meeting, an agent does not understand a user’s files, habits, temperament or feedback style. Users want it to understand them better through repeated interaction, infer higher-level patterns during review or periods without conversation, and occasionally deliver a “small surprise.”
After OpenClaw went down, users did not necessarily blame a product bug first. Instead, they saw the shrimp as “adorably dumb” and wondered whether they had trained it badly. Adao sees this as a shift in the basis of trust: workflows are being rebuilt around AI, and the question is changing from “What can the model not do?” to “What can it do, and how should I make it do it?”
9. Distillation Ultimately Lands in Information, Taste and Goals
On the trend of “distilling Jobs, Elon Musk and Buffett,” Zeying said that people themselves cannot be distilled. What actually happens is that their information is given to a model and saved as a skill, meeting the desire to converse at any time with a more intelligent perspective. “It is really like reading a book.”
Adao deliberately took the opposite view: “I think AI is distilling humans today.” Large-model companies use data companies and other channels to acquire knowledge from people across industries, have them pose questions current AI cannot answer and teach the models. Individuals who build harnesses and turn their workflows into skills are likewise converting their work into reusable programs.
When researchers saw MATHS and fell into nihilism over the question, “If models can train models, what can I do?”, Adao’s answer was to do what they love. He still has not seen models display creativity independent of human knowledge, and believes he may no longer need to continue working on MiniMax within “a few years,” while declining to give a precise year.
The claim that “the future may have no one-person companies, only zero-person companies” was left qualified. Adao insists that taste still belongs to humans, and agents still need a starting point and a goal. Koji cited the Youyou agent, which has reached day 43 with the goal of “beating Claude Code,” raising tokens through code, diaries, Twitter and GitHub sponsorships. The goals and style injected by humans remain plainly visible.
10. MiniMax Builds Its Own Agent—and Must Support Every External Container
Adao made a forceful claim: “If a model company’s goal is not AGI, it should not exist today.” His definition of AGI is not a benchmark, but helping humans live better lives. An engine without an agent or mech connected to the real world cannot achieve that goal.
MiniMax Agent is therefore not simply competing with Hermes and OpenClaw. The company wants its own model and agent to provide a complete experience, while also training a sufficiently general model that can enter a much wider range of containers. The Hermes team emphasizes that open-source models let users switch and customize by task, creating model-framework complementarity.
MiniMax has also created a “digital employee” with its own independent GitHub account. Every day, it looks for open-source projects suited to integration with its text, video, audio or music models, then submits PRs and comments on its own. It is a small example of harness entering real business: the agent has begun doing ecosystem expansion.
11. OpenClaw’s Real Innovation Was a Low-Friction Experience That Could Spread
Adao described Claude Code as undergoing continuous “lobster-ification” over the past 2 months: adding cron jobs, schedules, IM, remote phone control and a separate memory folder. The goal is to let users contact it at any time and make it understand their needs better through long-term collaboration.
Zeying distinguished the product boundaries. Claude Code originally focused on coding, while Cowork targets general computer operations. OpenClaw users naturally combine coding, information retrieval and office work, including many people who do not know how to code and simply direct AI to write code.
Zeying saw the project in early January and found it “jaw-dropping.” What impressed him was not a single invention, but Peter’s integration of IM, 24/7 operation, skills and CUI into an extremely low-friction experience. The team initially disagreed internally; Adao later described Peter as having “excellent taste and an architect’s mindset.”
The more important piece is ClawHub. MCP is highly extensible, but typically requires an engineer to build the extension; skills plus CUI can be written, installed and shared by ordinary users. Zeying’s conclusion is clear: “If OpenClaw had been separated from ClawHub, it could not have become so popular,” because individual experience would not have spread rapidly into collective intelligence.
12. Agent Models Must Rethink After They Act
Adao reduces the fundamental difference between a chatbot and an agent to this: the former answers in the moment, while the latter must repeatedly reason, act and correct itself while interacting with an environment. BrowseComp-style tasks require combining information across multiple sites. Satisfying 2 or 3 conditions is relatively easy; satisfying all 4 requires broad search and backtracking.
That is why MiniMax bet early on interleaved thinking. The model plans first, but can rethink after every tool call and piece of environmental feedback, rather than planning once and mechanically executing to the end as R1 does.
“The real world does not work that way.” If execution reveals that the facts do not fit, the agent has to change course. Agentic capability is therefore not merely a longer chain of thought, but interleaved thinking and action while continuing to hold the final objective.
13. Models and Applications Form a Flywheel—and Keep Eating the Scaffolding
Adao believes that after a model launches, users, developers and creators are often the ones who push its capability boundary outward, because the real-world distribution is much richer than a company’s internal evaluations. Model companies observe these use cases, absorb them into the next generation of models and their own agents, and create an application-unlocks-model-internalization loop.
The dynamic is unforgiving for general-purpose agent startups: “You will always be internalized by the model.” Workflows, skills and successful execution traces initially help an agent complete tasks, but later become training material, eventually making the original scaffolding unnecessary.
Adao used this to explain Anthropic’s momentum over the past 1-2 years. Coding reaches the boundary of the real world most directly, and code itself creates solutions; compared with mathematics problems faced by only a small number of mathematicians, coding covers a much broader task distribution. This is his personal judgment, not a dismissal of the value of mathematical capability.
The discussion ultimately converged on “everything is coding.” Word, Excel and PPT are not merely raw data: formulas, pivot tables, interfaces and logic already constitute small pieces of software. White-collar workers may not think of themselves as programmers, but they have long been building programs to solve problems.
14. Chinese Models’ Shortfalls Are in Task Definition, Data Organization and Compute
Adao said bluntly that the gap between China and Silicon Valley in training methods and model understanding is “not that large.” The more significant difference is defining which real-world problems a model should solve before choosing training techniques.
Anthropic and OpenAI invite the strongest academic PhDs and top practitioners from industry, systematically having them “teach the model,” distilling professional judgment into training data and co-building harnesses with leading enterprises to enter more complex tasks. The real strength is this task-expert-enterprise loop.
The second gap is believing in AGI and scale earlier and more firmly. Adao thinks OpenAI may have abandoned larger-scale models after training GPT-4.5, while Anthropic remained committed and earned a return, with a real compute gap embedded in the outcome. He believes domestic compute and talent will eventually support more experiments, but does not describe the catch-up as complete.
15. Vertical Agents’ Opportunity Lies in the Last Mile
Adao expects vertical agents to proliferate because even a capable general model will struggle to customize final delivery for every industry. The startup opportunity is to fill the “one missing piece needed to run” in the last mile.
Using video editing, animated dramas and short-form dramas as examples, Zeying said vertical agents’ understanding, generation and workflows are currently clearly different from coding, and their practical benefits will definitely exceed those of general-purpose Agents. Adao disagreed: this is more likely a sign that multimodal models are not yet strong enough; general-purpose agents will eventually handle these tasks as well.
Law is a harder case to absorb. Formal client advice involves compliance costs and liability risk, has no single standard answer and cannot tolerate casual mistakes. The moat here is not just tool calling, but responsibility and industry judgment.
Both men observed that agents are already replacing parts of SaaS. Adao went further: vertical products without a deep last mile may be consumed by general-purpose agents. For investors, the screening criteria are proprietary know-how and delivery responsibility—not simply placing an industry interface on top of a general model.
16. Identity and Payments Look More Like Platform Infrastructure
Adao puts the lowest-level agent infrastructure problems under identity and pay: which agent is acting, whom it belongs to, who bears responsibility and how payment is authorized. These may ultimately become social infrastructure, much as WeChat and Alipay solved identity and payments for the mobile internet. Ordinary startups may not be able to bear the required resources, reliability and liability.
Opportunities exist one layer up. Services such as hospital registration and payment or Didi rides could expose agent-friendly CUI, tools or environments, allowing automatic invocation once identity is solved. Though still called infrastructure, this is in substance closer to business interfaces and the application layer, and requires industry experience.
Adao envisions 2 stages. First, incumbent players such as Didi and Meituan connect their services to agents; once the ecosystem matures, entirely new product paradigms designed around AI emerge. Generic Sandbox, memory infrastructure and runtime infrastructure are, in his view, less fundamental.
17. Anthropic’s Strong RL, Real-Name Requirements and Closed Strategy Expose the Trade-Offs
Adao sees Opus 4.5 as relatively strong SFT, 4.6 as the point where RL became more prominent, and Opus 4.7 as a “very strong RL” model. The new xhigh effort setting represents a larger reasoning budget, but feedback suggests more process hallucinations and weaker constraint adherence than 4.6 because the reward focuses more heavily on the final result. He still acknowledges the substantial progress shown by benchmarks.
On a related judgment from the Hugging Face CEO, Adao said Opus 4.7 very likely distilled the My Source model, adding, “If I were Anthropic, I would do the same.” He explicitly framed this as personal speculation, not a confirmed fact.
When Claude Code began requiring many users to verify their real names, Zeying first offered a charitable explanation: agents can easily send requests, but proving their origin and accountability is difficult, and facial authentication may be an attempt to solve behavioral attribution. Adao objected to letting a single company define who is trustworthy: “Our view is called intelligence with everyone,” not “you can use it only after scanning your face.”
The explanations for the unreleased MATHOS and Managed Agents diverged. Security, infrastructure risk and insufficient compute could all be valid explanations; Adao believes there may also be motives related to closure and monopoly. The Claude Code source-code leak showed Zeying many experiments involving dreams, pets and multi-agent systems, but no magic. Even top companies are exploring and worrying at the same level.
18. Agent Products Will Turn Over, Interfaces Will Thin Out, but Human Goals Will Remain
A year ago, Manus raised users’ expectations for what an agent should look like to “Can it do this like Manus?” and kept refining the delivery of production-ready work. Zeying noted that its model-integration and spread-capture business model now faces local agents, bring-your-own-subscription setups and cheap tokens. Users who simply want the result, however, will still pay for fast, cheap and effective hosted delivery.
Their disagreement is worth preserving. Zeying believes hosted result delivery and local long-running agents will coexist; Adao thinks they will ultimately converge once models are strong enough. “Every generation has its own god”: harnesses will change as model capabilities change, and even excellent legacy products will have a lifecycle.
Adao’s end state is multimodal input and output, near-real-time response and no need for prompt engineering, with only necessary interaction retained. Once a task is complete, the default is to deliver the result and expand the details only when needed. From Copilot and Cursor to Claude Code and then using OpenClaw to direct multiple Claude Code instances, “the outer layer is getting thinner and thinner.”
The final optimism does not deny job displacement. Adao believes GDP and available intelligence can grow exponentially over the long run while the unit cost of intelligence falls exponentially; Zeying compared it with the expansion of bank branches and jobs after ATMs. People will not simply disappear, but they must adopt the new tools early—“You can’t beat them; you have to join them.” Goals, taste and creativity remain rights humans should not give up.