A policy researcher with 15 years of experience used open-source models and an Agent framework (programs that let AI autonomously plan and execute multi-step tasks) on consumer-grade hardware. Spending 5 hours and 6 autonomous iterations, they generated a 21-page report on the current state of AI in Europe — AI deep research has officially moved from "is it possible" to "usable but don't rush it."
What this is
This self-proclaimed unemployed public sector researcher used Hermes Agent (an open-source AI agent framework) paired with the quantized Qwen 3.6-35b (Alibaba's open-source LLM, compressed for local execution) to run a research report generation workflow on their own PC: starting from a draft, diagnosing issues, revising content, generating charts, and inserting them into the document, the six-round loop was completed almost autonomously. Even the code repository's README and folder structure were generated by the AI. The author open-sourced all prompts, scripts, and intermediate artifacts, explicitly stating this is a "starter scaffold" for peers, not a finished solution.
Industry view
We note that the value of this case lies not in the report quality itself — the author self-evaluated it as "not excellent, but good enough as a starting point." The real signal is that traditional knowledge workers without technical backgrounds have already closed the loop on Agent workflows using consumer hardware, and believe full automation is achievable. This serves as a warning for commercial deep research products like Perplexity — the barrier for users to build their own alternatives is dropping.
But the risks are equally apparent. First is speed: an RTX 4060 with 32GB RAM outputs only 28 tokens per second, taking 5 hours for one report; from a productivity standpoint, it's not cost-effective. Second is quality control: 6 rounds of autonomous iteration mean the AI is "self-reflecting," but without intervention nodes from domain experts, the accuracy and compliance of professional reports cannot be guaranteed. Critics will point out that a policy brief "looking the part format-wise" and "having correct content" are two different things — the hallucination problem in current models will only be amplified, not converged, during autonomous long-form generation.
Impact on regular people
For enterprise IT: The combination of open-source Agents and local deployment makes deep research with zero data exfiltration possible, but the 5-hour compute cost and lack of quality control processes mean it remains an experiment, not a production tool, in the short term.
For individual careers: The "draft generation" phase of traditional research and consulting roles is being eroded — it's not AI replacing you, but people who know how to use AI replacing those who don't; this judgment now has another concrete instance.
For the consumer market: Running Agents on consumer GPUs still leans toward a geek toy, but with model miniaturization and framework maturation, it could enter the daily routines of a broader range of knowledge workers in 12-18 months.