Building an AI Research System for a One-Person Firm: Progressive Determinism Architecture
otionflow · reddit · 2026-10-03
A Reddit user shares the architecture behind Dewey, an AI-assisted research/knowledge-work system for a one-person research business, and asks others what they're building.
Core idea: rather than one giant agent, it's a collection of systems around a strong generalist model (mostly Codex) — persistent project state, research/evidence workflows, provenance, reusable methods, deterministic checks, recovery/repair, and governance over what Codex works on and is authorized to do.
Key principle — progressive determinism: let flexible models handle messy work, observe what repeatedly works, then turn stable pieces into normal software, validators, or tightly bounded workflows. The flexible version stays available for cases outside the automated niche.
Separating work from worker: ideally the same bounded operation could be handled by Codex, a cheaper model, a specialist agent, a script, or a human, without the project's state and rules living inside that worker's context.
Flow: persistent state → bounded task/context → appropriate worker → structured result → validation → state update → next work.
The metric the author cares about most isn't token cost but human attention per unit of valuable output. He also asks which parts people initially thought needed agents but eventually turned into normal software.
More from AGI Musings
- RIP traditional data scientists: the role is splitting into two tracks — mdancho84 · 2026-10-03
- David Patterson: solar's near-vertical cost curve will power the singularity — davidpattersonx · 2026-10-03
- ryolu_: When everyone copies what works, we stop making anything of our own — ryolu_ · 2026-10-03
- Yann LeCun at ETH Zürich: Scaling LLMs to reach AGI is 'impossible' — rohanpaul_ai · 2026-10-03
- VC accuses OpenAI of skynet-style regulatory capture, says Anthropic's version is more subtle — StewartalsopIII · 2026-10-03
- Peking University mathematician: AI's millennium-prize wins are 'harvesting,' not discovery — 新智元 · 2026-10-03