A Retrospective on the Evolution of 2C and 2B Agents | A Conversation with Afk.surf Founder 贺嘉琛
Topics
Autonomous Agents Reshaping Enterprise SaaS
Personal AI Assistants Adoption and Monetization
How AI Reshapes Enterprise Workflows
AI Inference Cost Optimization
Specialized SLMs vs General LLMs
Enterprise AI ROI and Cost Optimization
Open Source vs Closed Source AI
Pioneers
贺嘉琛
Summary
- The agent product form may be stabilizing, with competition shifting toward the ability to take over complex systems. 贺嘉琛’s move from AFFiNE to AFK reflects his view that as models become more autonomous, specific editing environments can become a constraint. Agents are no longer tied to a particular interface; they can respond anywhere via text, phone, or office software. The next frontier is deeper gains in intelligence and tool use.
- Personal agents for consumers are a platform opportunity, but startups will have to fight for supply chains and distribution. 嘉琛 argues that user numbers may not create an intelligence moat; model progress remains the biggest variable. The real differentiators may be purchasing power, pricing, and service supply. Platforms will resist becoming suppliers and may ban or retaliate against challengers; startups also have to subsidize goods, tokens, and advertising at the same time.
- Equipping everyone with AI does not necessarily make the company more efficient. A product manager might ask AI to expand requirements into a 300-page PRD, then an engineer might ask AI to compress it into one sentence—skipping the necessary conversation: “What gets amplified is more likely noise than signal.” AFK’s approach is to give everyone the same agent, with personal memory and an execution environment, and try to turn cached context, work experience, and business understanding into shared organizational capabilities.
- The certainty and cost advantages of workflows are being redefined by model capabilities and underlying reuse mechanisms. 曲凯 presses 嘉琛 on why enterprises adopted workflows in the first place: to ensure closed-loop execution and lower costs. 嘉琛 says those capabilities are moving “from outside the model to inside it.” He sees instruction following, reinforcement learning, KV cache, prompt cache, and distillation as ways to replace some external orchestration, while warning that old skills and system prompts may start to have side effects.
- The opportunity for startups training small models lies in reproducing tasks large models already do well, not in pushing the capability ceiling higher. Once skeptical because of the compute, talent, and infrastructure gap, 嘉琛 has shifted to “muscle memory”: capture task traces from leading models, then distill them around user preferences to handle repetitive work at lower cost and with faster response times. AFK is also training a task routing model and an approximately 20–40B-parameter work model for internal enterprise operations.
- Enterprise adoption of agents depends on clear returns and a willingness to provide business context. 嘉琛 believes companies will actively support system changes when leaders see that work previously done by 8 people can be completed end to end, and can trace how tokens flow into code, testing, and deployment. That is also why a small team chose 2B: “We’d rather work on things closer to customer value.”
- AGI model supply may become abundant over the next 2–3 years, making today’s model advantages less scarce over time. 嘉琛 expects open-source models to catch up faster and sees an immediate opportunity to help already profitable companies capture more value. Further out, dynamically generated software could reshape applications and business rules. 曲凯 argues that agents have only recently started approaching the commercial maturity of companionship and video; 嘉琛 agrees with the premise: value is unlocked only when the agent completes tasks end to end.
Deep dive
1. AFFiNE caught the multimodal wave, then ran into the limits of the editing environment
- 嘉琛 credits AFFiNE’s early traction to timing: work already spans documents, recordings, canvases, and data processing, so those tools belong together. When the team raised funding in 2022–2023, it was already pitching a “multimodal workspace.” He says the project eventually became roughly the leading open-source product in its category.
- The team trained its own embedding model and multimodal models. In 嘉琛’s account, when AFFiNE AI launched in April 2024, it was the first workspace able to retrieve across multimodal information, conduct Deep Research, and generate content. Its positioning was like “a canvas for people who don’t code,” for working with documents, spreadsheets, and design files.
- But in the second half of the race, he came to see specific editing tools as a “constraint” on models moving closer to AGI. From Cursor to Claude Code and then agents everywhere, products are changing as autonomy rises. AFFiNE went through a similar shift internally in 2024–2025; changes in capital markets also prompted him to start again.
2. Agents are breaking free of fixed interfaces; the underlying challenge is orchestrating complex systems
- 曲凯 asks whether personal agents moving into the mainstream means this product evolution is nearing its end. 嘉琛 thinks the product form “may already be fairly settled”: agents are increasingly like people you can reach through any channel. Further gains will come mainly from intelligence and tool use.
- 嘉琛 sees Poke as an important precedent. From the second half of 2025 into the first half of 2026, users could text it over iMessage to get things done, and it could also reach out to them. “Always on, always available” is closer to the shared goal for this generation of products than having a standalone app.
- The underlying capability is evolving from calling a single API, to managing complex API systems, to coordinating tools across multiple operating systems. Browser Use, Computer Use, CLI, and MCP are just means to that end. The shared technical challenge for 2C and 2B is to let AI operate complex systems across multiple endpoints and for multiple users.
3. 2C must rebuild consumer platforms; 2B offers a more direct path to value capture
- 嘉琛 is cautious about the moat in 2C: more users and behavioral data may not create the same economies of scale as in earlier internet products. A chatbot does not naturally evolve into an agent just because it has more users. Enterprises are different: people working toward a shared goal create room for collective learning through repetitive work and information loss.
- He sees products such as Instinct and Muse in the US as trying to reinvent platform commerce and rebuild the consumer supply chain. Whether they can save users money, get them what they want, and withstand bans from food delivery and travel platforms may matter more to the outcome than the agent product itself.
- 曲凯 suggests that if startups succeed, lifestyle platforms such as Didi and Meituan could retreat into the background as suppliers. 嘉琛 cautions that those companies remain powerful and will compete directly. “No platform wants to go from owning the interface to being a supplier.” Companies once subsidized only goods; now they also have to subsidize tokens and advertising.
- For a small team, the 2B value chain is more direct: it is easier to see what a dollar of tokens taught the system, what it produced, and whether it reduced costs. 嘉琛 chose the enterprise market because it lets the company capture customer value sooner, rather than first bearing the cost of rebuilding an entire consumer platform.
4. Even after everyone starts coding with AI, organizational collaboration can remain the bottleneck
- 嘉琛 attributes winning major customers to the initial wedge. Consumers may not be willing to hand over all their context for recommendations and shopping, while enterprise leaders have an incentive to help change systems when they see work previously done by 8 people completed end to end. He wants to connect business and finance processes to the path each token takes into code, through testing, and into deployment.
- 曲凯 asks what AFK adds when internet companies already use AI coding widely. 嘉琛 points out that work outside engineering involves GUIs, more environments, and more data, making it harder. As AI starts delivering finished work, designers can ship production features, videos, or onboarding; product managers become more like engineers who define feature logic.
- The sharpest counterexample is a 300-page PRD: a product manager asks AI to expand the requirements, then an engineer asks AI to summarize them in one sentence before writing code. The discussion between people that actually matters gets skipped. He says some companies saw productivity fall after giving everyone Max plans because “what gets amplified is more likely noise than signal” (“放大的大概率是噪声,而不是信号”).
- 嘉琛 argues that product managers, designers, and engineers should work together in a code-based monorepo, using artifacts to build, preview, and revise results in real time. As AI moves from “typing for you” to “making decisions for you,” people’s roles and workflows also have to change.
5. Workflow certainty and reuse are moving inside the model
- 嘉琛 observes that n8n was discussed much less by late 2025, while individual skills that drew heavy attention in April–June 2026 also quickly lost momentum by the second half of the year. His explanation: models may not need better workflows; they are starting to need the old workflows less. “The models have learned them.”
- That changes the purpose of engineering work, too. Many harnesses and developer platforms used to compensate for AI’s shortcomings. After switching to a smarter model or a better harness, the first step may instead be to remove old skills and delete system prompts, because the original instructions can introduce side effects.
- 曲凯 raises the enterprise user’s counterpoint: workflows were built to ensure closed-loop execution and reduce the cost of repetition. 嘉琛 responds that models used to hallucinate more and follow instructions less reliably, which made fixed external outputs necessary. He believes advances in reinforcement learning and instruction following have made many API operations more reliable.
- Cost reuse does not have to depend entirely on explicit orchestration. 嘉琛 says agentic workloads have high KV cache and prompt cache hit rates, and calls worth repeating can also be distilled. He sums up the shift as workflow logic moving “from outside the model to inside it.”
6. One enterprise agent aims to turn individual experience into organizational capability
- AFK’s approach is to “give everyone the same agent,” while supporting high concurrency and personalization. Each person still has their own memory and execution environment in private conversations, but the system learns continuously across employees, software, and business processes instead of leaving people with disconnected personal assistants.
- 嘉琛 uses a hypothetical example to explain the value of caching: 100 people using separate agents might have to read the same code 100 times; a shared system could reuse context a colleague has already read. More important to him is letting other employees benefit from the effective practices developed by the 2 or 3 people at a company who are best at using AI.
- He wants the agent to see code, APIs, meetings, and customer conversations together, identify business bottlenecks, and use parameterized training to give the company a continuously learning system. Otherwise, a boss with 100 employees and 100 personal agents may still end up with “200 different answers.”
7. Small models do not need to beat SOTA; the business case for training them has changed
- In the second half of 2025, 嘉琛 took part in training the first multimodal agentic model in China with more than 1T parameters. That experience made him “especially skeptical of the trend of startups training models”: large-scale RL requires talent, compute, and scientifically generated data, and frontier labs have much stronger end-to-end capabilities.
- In the past, launching a model often meant reaching SOTA on certain benchmarks, which encouraged overfitting and produced short-lived leaderboard advantages. He sees a new opportunity in 2026: with lower training costs, changes in architecture, and foundation models reaching the required capability level, small models can reproduce what large models already do well without raising the capability ceiling.
- The goal of this “big model teaches small model” approach is to turn long reasoning processes into System 1 behavior that no longer requires repeated deliberation. 嘉琛 says the commercial case depends on 2 conditions: user preferences that serve as a reward signal, and correct results from a strong SOTA model as ground truth. With both in place, lower cost and latency have practical value.
8. AFK is training muscle memory and the ability to operate enterprise systems
- The first category is a task routing model: record tool-use traces from advanced models as they complete tasks, then distill them into a smaller model. 嘉琛 gives the example of a task with 5 viable paths, where a user may prefer 1 or 2 of them. The small model has to learn both “how to do it” and “what the user prefers.”
- He compares it to learning to ride a bike or swim: the first attempt takes a lot of thought; practice turns it into a reflex. A smart large model is like the cerebral cortex; the small model provides the “brainstem and spinal cord,” creating low-cost “muscle memory.” 曲凯 connects this to DeepSeek’s division between the brain and cerebellum. 嘉琛 cautions that DeepSeek’s approach to that division is also being overturned now.
- The second category is a work model with roughly 20–40B parameters, designed for internal enterprise tasks. 嘉琛 says closed-source models still face safety restrictions when operating certain systems or dynamic software. For example, when a company needs to reverse engineer its own systems or interfaces, AFK aims to provide stronger capabilities with fewer task restrictions.
9. Dynamic software and more abundant model supply could reshape the next competitive cycle
- 嘉琛 gives an example of a skill that dynamically writes instruction sets and rendering pipelines for different Switch games and Android devices so the games can run on those devices. From that, he imagines software tailored to each person and each task, with the boundaries of general-purpose software continuing to shrink.
- His more provocative claim is: “All software today is open-source software.” He does not mean the source code is actually public; rather, if AI knows what a piece of software does and has its API spec, it can reverse engineer or rewrite the same functionality. More and more software logic could therefore shift into model intelligence.
- On timing, he expects AGI model supply may become abundant in 2–3 years, with open-source models catching up faster. For now, the strongest models still largely come from Anthropic and OpenAI. AFK can first help companies that are already profitable capture more value, and earn a profit in the process.
- 曲凯 asks whether agents are only now approaching commercial maturity, given that companionship and video have already achieved viable business models. 嘉琛 agrees that end-to-end delivery is key. He points to Opus 5.5 editing video and GPT-6 Astra’s world-model capability for controlling robots as examples of general models challenging categories that matured earlier: “This bucket is starting to fill up.”
10. The return from a startup also includes giving its people a stronger shot next time
- Asked about talent and organization, 嘉琛 offers no specific hiring playbook. He cares more about giving talented, thoughtful people fewer constraints than they would have at a big company, and working with “smart, kind, persistent, and brave people.” Even if not every company goes public, people can still build skills, increase their value, and improve their odds of succeeding next time.
- He points to AFFiNE’s former CTO 太郎, who later became a Principal Engineer at Cloudflare, and to his own experience helping build Void Zero from its founding through its acquisition by Cloudflare and earning an exit return. “Growth is the greatest return from entrepreneurship,” and “Shoot for the Moon” may not guarantee you reach the goal, but it can help you become more than you were.