I recently pulled down and read through both EvoScientist and OmniScientist. Both ship with their own arXiv technical reports and both claim end-to-end research capability. But reading the source code, you quickly find they are not solving the same problem at all. One is trying to make AI act more and more like a seasoned researcher; the other is trying to stop AI from hallucinating over second-hand spreadsheets.
Below is a comparison based on the two READMEs and source code — not an official write-up, and it includes some of my own judgment.
EvoScientist: A Self-Evolving Research Partner
EvoScientist (v0.2.8, Apache 2.0) is built on LangChain's DeepAgents framework. Its core selling point is "self-evolution": after each research run, the system records observations from the process into a knowledge graph, with edges labeled complementary / contradictory / superseded. Even more aggressive is the AutoSkills mechanism, which periodically mines its own memory for recurring patterns, automatically drafts them into reusable skills, and holds them for review via /autoskills. In other words, the system writes its own tools.
Alongside this are 6 sub-agents: planning, research, coding, debugging, data analysis, and writing, plus an orchestrator. The human sits "on the loop" (Human-on-the-Loop) rather than "in the loop": by default every shell execution requires your approval, and agents proactively ask when uncertain. It integrates Telegram, Lark, WeChat, Slack and more, with WebUI, TUI, and CLI, and even supports cron-style scheduled tasks via /schedule. It's remarkably mature — it ranked first in two versions of the DeepResearch Bench and won Best Paper at ICAIS 2025.
OmniScientist: A Referee-Style System That Reads Raw Data
OmniScientist (v0.1.1, MIT) is obsessed with something completely different. Its README opens by noting that most agents today reason only over text and code — raw observations like microscopy images, seismic waveforms, and CT volumes must first be compressed into tables by humans before being fed in. Its goal is to let agents read the raw materials directly.
Implementation-wise, it is a self-contained ReAct loop with three phases: Ideation (hypotheses) → Experiment (run code, read images) → Writeup (write the paper), with a "gate" blocking the exit of each phase. The gate is code-based and reads execution records, not the draft: every number appearing in the manuscript must have actually shown up in some run's stdout, or it doesn't pass. Empty experiment results bounce back to the ideation phase, with up to two rollbacks — and it remembers the previous failed line of thinking. References are verified in real time against OpenAlex and Crossref; if it can't verify, it won't fabricate — unlike many "AI writes papers" projects.
It supports 12 modalities spanning images, waveforms, audio, video, point clouds, trajectories, tables, and formulas, with sample cases in seismology, chemistry, genomics, astronomy, pathology, CT, superconductivity, bird sounds, and more. The desktop app extracts credentials from the environment before handing code execution to the agent — a detail that shows real security awareness.
How Similar Are They, Really?
A quick checklist:
| Dimension | EvoScientist | OmniScientist | |---|---|---| | Primary goal | Continuous cross-session evolution | Raw-data perception with strict provenance | | Agent structure | 6 cooperating sub-agents | Single ReAct loop, three phases | | Multimodality | Weak, mostly text/code | Strong, 12 native modalities | | Verification | Human approval + ask_user | Code gates + stdout number back-linking | | Coverage | General research assistant | Multi-domain STEM samples | | Deployment | Python stack + WebUI | Desktop app / CLI, local-first | | Maturity | v0.2.8, benchmark first | v0.1.1, early stage | | License | Apache 2.0 | MIT |
They also share quite a lot: both produce an end-to-end paper with figures and tables; both insist the agent runs its own analysis code and reads back the results rather than memorizing; both put humans in a critical position — one via approvals, the other via gates; and both picked the most permissive licenses.
Their Respective Weaknesses
EvoScientist's "self-evolution" sounds impressive, but it works in a text/code-centric way. It cannot "see" an electron micrograph or a seismic waveform — raw observations still need you to convert them into something it can read. Its anti-hallucination mostly relies on humans: approvals, dangerous-mode flags, auto-allow lists — the system itself lacks OmniScientist's hard constraint that "numbers must come from real stdout." Also, the full stack depends on a langgraph dev backend, and the WebUI requires Node 24 — deployment isn't light.
OmniScientist, conversely, has strong verification and strong perception, but it's a single-task, single-run design with no cross-session memory or skill accumulation. Its ecosystem is much smaller — few built-in skills, no messaging channels, no scheduling. v0.1.1 means interfaces will still change, and there's no feedback yet from real Windows machines. If you expect it to "understand you better over time," that's not there yet.
My Take
These two are not competitors — they're two pieces of a puzzle.
If you want a long-term research partner that accumulates experience, builds its own tools, and can take tasks through messaging channels, EvoScientist is usable today and works smoothly. It bets that the process gets smoother with every run.
If you want to throw a specific dataset at it and get a report where every number is verifiably grounded in real data, OmniScientist's design fits better. It bets that "seeing reality, no fabrication" is more foundational than "evolution."
One personal concern: both are still stuck in a "building a god" narrative. EvoScientist makes self-evolution sound thrilling, but whether cross-session memory could become a self-reinforcing echo chamber has never been validated. However hard OmniScientist's gates are, the hypotheses and the direction are still set by humans. Before either can genuinely assist publishable research, there's still a stretch of engineering to cover in explainability and accountability.
That's it. Both repos are on GitHub — links and technical reports below. Reading the source yourself is faster than reading this post.
References
- EvoScientist: https://github.com/EvoScientist/EvoScientist — technical report arXiv:2603.08127
- OmniScientist: https://github.com/Omni-Scientist/OmniScientist — technical report arXiv:2608.13558