13 minutes spent hunting for a single earlier suggestion turned into an open-source plugin—this happened on DeepSeek Harness, with not a single line of manual code written by a human. What we think is worth watching: when the platform AI runs on becomes programmable by AI itself, the wall between regular users and "developers" is thinning.
What This Is
DeepSeek Harness (DSH) is a development platform under DeepSeek for running AI Agents. An Agent is an AI that autonomously calls tools and completes tasks in multiple steps—not the kind that only answers a single question. Benyi chatted with the AI inside DSH for over twenty rounds, trying to recover one earlier suggestion. After scrolling through 13 minutes of chat history, he gave up. He didn't file a ticket, didn't open an IDE (the software programmers write code in). Instead, he typed into the chat box where the AI was already working for him: build me a chat directory. The AI fired back three clarifying questions (sidebar or floating button, summary or full text, temporary test or permanent save), then got to work. It crashed once (a page error), came up empty once (an architecture issue), and on the third iteration it worked. He converted the temporary plugin into a permanent version and open-sourced it on Gitee (a domestic code-hosting platform). What he did the entire time: answered three multiple-choice questions and refreshed the page three times.
How the Industry Sees It
The legible signal here: an Agent modified the very platform it runs on. In the past that meant filing a ticket, waiting for a sprint slot, writing code, and shipping a release. Now it happens through conversation.
To stay clear-eyed, a few caveats to discount the headline:
This is a developer scenario, not an ordinary office workflow. DSH itself is a development platform built for Agents; everyday office software is still far from this capability.
The AI failed twice before getting it right—one page error, and one instance where ambiguous configuration semantics caused new entries to be silently skipped. Even for its own runtime platform, the AI doesn't fully understand the interfaces.
The semantic distinction between "modify" and "create" in the configuration was something the AI had to dig out on its own. These "system-level landmines" remain where Agents frequently crash and burn—you can't stabilize them with a clever prompt.
The whole story took place in a developer-familiar domain. Change the scenario and ask the AI to modify a company's finance system, and the outcome would almost certainly be very different—don't extrapolate this example to all software too quickly.
Impact on Regular People
For enterprise IT: worth tracking, but don't rush to deploy. Agent self-modifying platforms are still a developer-tool play; enterprise ERPs (systems managing finance, inventory, and HR) and CRMs (customer relationship management systems) aren't there yet—but we expect the first batch of "conversationally modifiable" internal tools to appear within two years.
For individual professionals: what's usable today is "conversational automation"—let AI take over your repetitive operational needs. You don't have to become a developer, but you do need to learn to articulate requirements clearly.
For the consumer market: mobile apps and office software are still far from this. We expect "fixing bugs and requesting features via chat" to spill over from the developer community to mainstream users over the next three years.