In the AI Era, How Do You Boost Your Own Productivity?
When execution cost approaches zero, the bottleneck shifts to human cognition. The rational move is to build your own boat that grows with AI, not to cling to today's strongest AI product.
When execution cost approaches zero, the bottleneck shifts to human cognition. The rational move is to build your own boat that grows with AI, not to cling to today's strongest AI product.
TL;DR
- AI coding ability has leaped forward in the past six months; execution cost is approaching zero. The real bottleneck is no longer code, but human cognition.
- The strategy is not to "pick the strongest AI product right now," but to build yourself a boat that grows stronger alongside AI.
- The boat must be "your own" — generic products cannot solve personal problems, and the cost of building your own has dropped low enough to finish over a weekend.
- This article records the first version of my boat: orchestration, information, access, knowledge accumulation, and what to do next.
The most striking impact came from comparing two Helm requirements.
Last June, when working on the Coze-Studio open source project, we needed to support Helm deployment. Knowing nothing about Helm, I used Gemini 2.5 Pro to learn while building, and it took 3-4 days to get a working Helm Chart.
In December, another Helm-related requirement came up. I just said one sentence to Claude Code Opus 4.5 — and it wrote the entire Helm Chart for me.
After that, I kept asking myself one question: in an era where AI is this powerful, what kind of person still has an advantage?
My answer: the core competitiveness of a worker in the AI era is to use AI as leverage — to multiply your own efficiency to the maximum.
Recently I was working on an ACP protocol project, letting an AI Agent run a super-long task — autonomous development, review, refactoring, running dozens of loop cycles overnight.
Then the problem came: I spent a full 1-2 days just to understand the code AI wrote, figure out how to adjust the architecture, and then have AI refactor it according to my ideas.
AI was no longer the bottleneck — the bottleneck was the depth of human cognition.
Thinking about it differently — if I had known the ACP protocol inside out from the start, and had thought through the architecture design clearly enough, I wouldn't have needed to wait for AI to generate code before reviewing. I could have just told it "write it this way" and let it execute.
This also explains why the AI era amplifies the advantage of "experienced people" — AI pushes the cost of "execution" to near zero, so the value of "thinking clearly" becomes higher than ever.
Since humans are the bottleneck and AI is getting stronger fast — the most rational choice is to build yourself a boat that can grow with AI.
The Claude Code product lead said something that stuck with me:
"Don't build products for today's model, build products for the model six months from now."
The Manus team said something similar:
"If the progress of model technology is a rising tide, we want Manus to be a boat on the water, not a pillar fixed to the seabed."
Both quotes say the same thing: AI is advancing rapidly, and you should build a boat, not a pillar.
Translating this idea to the "individual" level:
What you should do today is not pick "the strongest AI product right now," but build yourself "an Agent that grows with AI."
Once this boat is built, you'll always be on top of the wave.
You might ask: there are already plenty of ready-made Agent frameworks — OpenClaw, Hermes... why build your own?
My answer in one sentence: generic products cannot solve personal problems.
Everyone's workflow is different, and the tools they use are different. Different roles do completely different things, with completely different workflows and tools. Currently, generic Agents are still at a very early stage, and I believe that for a long time to come, generic products will not be able to cover the diverse needs of individuals.
Side note: I recently observed the market and found many AI products are chasing surface-level things — "context optimization," "product form upgrades"... but with model capability gaps already flattened, I think a product's real moat is not in these surface-layer things.
What really made me decide to "build my own" was something else — the biggest change in the AI era is that software development costs have dropped dramatically.
In the past, building an AI workbench might take several full-time engineers half a year. Now, just me plus Claude Code, working evenings and weekends, can get the framework up in two or three weeks. Since the cost of building my own has dropped this low, why not do it?
My Agent currently has five layers of capability, each corresponding to a specific use case:
Below, I'll pick one dimension per section and explain in detail.
| Layer | Capability | Scenario |
| Protocol | ACP connects multiple CLIs (Claude Code, CodeX, COCO, etc.) | Enable "one brain + many hands" |
| Orchestration | Multi-Agent recursive Review / Refactor | Let AI write, check, and fix its own code |
| Information | Scheduled tasks + active fetching | Tech trends, Oncall, release notifications auto-delivered |
| Access | Feishu / WeChat / mobile web | Code while lying in bed |
| Knowledge | WIKI + symlink shared context | Let AI and I share the same knowledge base |
The original motivation for this module was simple: Claude Code's built-in Ralph Loop didn't meet my needs. When I use models to review code, the real flow is "Review first, then Double Check, then Refactor, then Review again" — a recursive loop.
Code Review → Double Check → Code Refactor → Code Review → Double Check → Code Refactor → ...
If Code Refactor is also handled by the model, this process doesn't need human intervention. So I built a "recursive tree structure" loop capability:
Two key design decisions:
This pipeline works well for both Review and Refactor scenarios. An interesting byproduct: the ACP network transport link parameter configuration — only AI could consider all these details and produce a configuration that balances latency, timeout, and retry.
For scheduled task scenarios, I mainly have three needs:
The overall experience is quite good — new tech trends are automatically received, and it can also help me with summaries, so I don't need to manually refresh web pages (when busy, it's easy to miss messages).
Since the Agent can already run 24/7, accessing it from a phone anytime is a natural need. So I integrated Feishu and WeChat — but since neither supports streaming output, the experience isn't as good as using a VPN to connect directly.
But once you can access it from a phone, you really can code while lying in bed. Last weekend I literally lay in bed and let it finish a small feature — the experience was quite surreal.
I consolidated my knowledge base WIKI into a single git repository; then in each code repository, I use symlinks to link to the corresponding project folder in my personal knowledge base. This way, when AI works on related tasks, it can share the same knowledge context as me.
In most scenarios, after a dozen or so rounds of repeated Review + fixes, there are basically no bugs. But I have encountered some corner cases:
Anthropic, in "Scaling Managed Agents: Decoupling the brain from the hands," breaks an Agent into five components:
This is basically the same architecture I outlined when doing Coding App research.
Following this framework, here's what I want to do next:
The software engineer profession won't disappear, but the job structure will change dramatically. What's truly worth thinking about is not "will I be eliminated," but "how to seize the dividend of this transformation."
If you also want to build your own Agent, my advice is — start with the smallest closed loop: pick 1 repetitive task you do every day (writing weekly reports, tracking tech trends, organizing meeting notes...), use ACP / Claude Code to automate it, and slowly expand from there.
Let everyone have their own AI Agent, let individual productivity be infinitely amplified — this is both a gift from our era and a test left for us.
Finally, a quote from the end of the DeepSeek paper to encourage us all:
"Be not tempted by praise, not intimidated by criticism; follow the Way and act, upright and correct in oneself."
我也来评论一下吧
测试一下