Skip to content
Pero: Open-Source AI Agent Workspace for Scientific Research Workflows

Pero: Open-Source AI Agent Workspace for Scientific Research Workflows

The author introduces Pero, an open-source project applying AI agent capabilities from coding assistants to scientific research workflows, featuring project-centered workspaces, distributed agent collaboration, and domain-specific apps like Virtual Cell for cell biology, built with Python, Rust, Go, and Electron.

phodal
phodal
phodal

From AI Coding to AI for Science

The author, previously working on AI coding agents at Qoder, decided to pivot toward applying agent capabilities — task planning, tool use, long-term memory, multi-agent collaboration, and execution environment management — to scientific research workflows (AI4S). The goal is to move beyond one-off coding assistance and enable agents to participate in ongoing research projects, organize literature, data, experiments, and coordinate distributed compute resources.

Why AI for Science (AI4S)?

Research workflows often start with a Python script, notebook, or model demo and evolve to involve literature, data, experiments, compute resources, and result validation. These stages resemble software engineering but demand stricter reproducibility: data provenance, method and environment tracking, and independent verification. Pero is an open-source workspace that connects domain-specific apps, agents, and compute resources, starting with Virtual Cell as a concrete entry point.

Pero concept diagram
Pero concept diagram

Three Entry Points for Pero

1. Project-Centered Workspace

Instead of single-turn chat, Pero organizes research questions, literature, data, tasks, experimental results, and knowledge around a persistent Project . Agents understand project context, track progress over time, and proactively suggest new tasks when data or conditions change. The aim is continuous research collaboration with full project history, long-term memory, and proactive workflows.

Project-centered workspace UI
Project-centered workspace UI

2. Open Distributed Agent Collaboration

Science requires diverse expertise and tools. Pero supports multiple agents dividing work within a project and dispatching tasks to authorized local machines, lab servers, or cloud resources. The vision extends open-source contribution beyond code: contributors may provide agents, tokens, compute, and experimental results. A basic implementation of Task, Run, and Worker exists, with GitHub task import, claim, and distribution. A task board is available at https://app.openpero.com/tasks (server performance noted as slow).

Task board screenshot
Task board screenshot

Full multi-agent research collaboration remains distant; automated contribution, artifact registration, and independent verification need further work.

3. Domain-Driven Vertical Scenarios

Different scientific domains have unique data formats, computational models, and validation methods. A generic chat interface cannot capture these differences. Pero adopts a composable research app model: domain capabilities are exposed via dedicated interfaces and can later be invoked as tools by agents in complex workflows. Examples include:

Life sciences: cell atlases, gene expression, protein structure analysis.

Materials science: molecular modeling, crystal structures, physics simulations.

Concrete Example: Virtual Cell

To validate the direction, the author built a Virtual Cell prototype based on a former colleague's question. Using ChatGPT for literature research, they created an interactive app with:

Reference cell atlas browsing

Perturbation analysis pipeline

First-order reaction simulation

3D volumetric data visualization

Users can inspect cell data, adjust parameters, and observe outputs without switching between scripts and tools. The author emphasizes that reaction simulations and volumetric examples are interaction demos, not scientifically validated predictive models. Perturbation analysis has limited data and evaluation scope and cannot be generalized to reliable predictions for unknown conditions.

Virtual Cell interface
Virtual Cell interface

Virtual Cell continues earlier experiments from the Point project and is now integrated as Pero's first domain app. The local UI and Python backend run; agent tool calling, Task/Run execution, and Project artifact registration are not yet connected.

The author envisions a workflow where a researcher poses a gene perturbation question, the agent assists with literature review, method selection, and experiment planning; upon confirmation, the domain app or remote compute executes the analysis, and inputs, method versions, execution environment, and results are saved to the Project for later audit and iteration.

Technology Stack

Pero uses a polyglot stack aligned with workspace, coordination, and domain compute concerns:

Python : scientific computing, data processing, domain app backends; integrates methods and models per scenario.

Rust : GPUI-based workspace, local host, ACP integration, standalone workers; current desktop path uses GPUI/Wasm, with native desktop exploration underway.

Go : cloud task coordination and distribution; custom Agent Harness is a future direction.

Electron + JavaScript/TypeScript : desktop shell, domain apps, visualization, web integration.

Design references include Zed/GPUI (workspace and UI), Kiro Crew (persistent workspace and app composition), Claude Science (research workspace), and DeepSeek Harness (plugin architecture). A harness PR is at https://github.com/phodal/pero/pull/4.

Current implementations: macOS desktop dev build, ACP sessions, distributed task basics, and Virtual Cell. The end-to-end flow linking domain apps, agents, and research result verification remains to be completed.

Summary and Call for Collaboration

The shift from AI coding to AI4S represents both a directional change and a relearning process. The author hopes to validate ideas through open-source collaboration with researchers, AI practitioners, and developers. If you have a repetitive manual research workflow, you are invited to share concrete details: inputs, current process, and result evaluation criteria. Contributions of domain apps, analysis methods, or review of Virtual Cell's methods and examples are welcome.

The final form of Pero is uncertain; the approach is to dig the pit, explore by building, and invite others to join. GitHub repository: https://github.com/phodal/pero.

Original Source

Signed-in readers can open the original source through BestHub's protected redirect.

Sign in to view source
Republication Notice

This article has been distilled and summarized from source material, then republished for learning and reference. If you believe it infringes your rights, please contactadmin@besthub.devand we will review it promptly.

AI Agentsopen-sourcemulti-agent systemsScientific ComputingAI4Sresearch workflowsVirtual Cell
Reader feedback

How this landed with the community

Sign in to like

Rate this article

Was this worth your time?

Sign in to rate
Discussion

0 Comments

Thoughtful readers leave field notes, pushback, and hard-won operational detail here.