Autonomous AI content matrix, 30 days in: context pollution, not prompts, was the bottleneck
Long-running AI workflows fail differently from short ones, and this run shows the fix is cheap and mechanical: isolate each run's working directory and inject a local knowledge base. Any developer building autonomous agents, batch content systems, or unattended pipelines will hit the same context-amnesia failure the moment their input changes between runs.
WorkBuddy schedules and orchestrates the whole pipeline, while DeepSeek Harness handles local file access, Markdown reading and writing, and Python script execution directly in a mounted working directory. Each morning the system scrapes GitHub Trending and Hacker News, analyzes source code for a featured project, generates a technical analysis with an architecture diagram from a locally fine-tuned Flux model, and queues it for a one-tap approval. On a 16GB RAM desktop it ran for 30 days, produced 4,000+ followers across platforms, placed two ads, and cleared 2,100 yuan in net profit.
Within a week the pipeline started failing in a familiar way. When the daily scrape switched from a Vue project to React, the AI wrote Vue's reactivity explanation into React's code walkthrough, and several consecutive posts opened with the identical boilerplate line 'With the continuous development of front-end technology...'. The diagnosis is that long-running agents accumulate context pollution: the model stops distinguishing current local context from historical memory. The fix was to give each day its own timestamped working directory, physically separating that day's sources from any prior data, and to add a lightweight local RAG that injects a term-calibration dictionary from SQLite into the system prompt via vector similarity search.
The build itself used little hand-written code: the author estimates 80% of it came from Cursor autocomplete, leaving 20% as wiring and failure-proofing between AI tools. That split points to a shift in what builders actually do: define the problem boundary, design fault tolerance, and assemble cheap toolchains into a cash-flowing product. The pure code-translator role is fading; the system architect who controls data flow and expects hallucinations is the one who gets stable output.
Context pollution in long-running agents is the practical barrier to autonomy, and the timestamped-directory plus local-injection fix treats it as a systems problem, scalable with more engineering rather than more prompt craft.
The author's decision to keep a local SQLite dictionary instead of live web search trades freshness for controllability, a defensible trade-off for publishing pipelines where a hallucinated technical fact is costly.
The 2,100 yuan profit is small, but the salient point is the operating cost: a single operator supervising a system that runs unattended on a modest machine, with 80% of code generated by an assistant. The bottleneck shifts from writing code to designing the supervision and failure-handling layer.