AI Agents for Community Scaling¶
Status: In development
The AI agents described here (N2-Whisperer, Mind-of-a-Worm, Mad-Worm-Scientist) are the model OpenWorm is building toward. First components are expected in Q2 2026. Today, contributor onboarding still goes through the volunteer application form — see the Community page for the current path. This page describes the target state.
OpenWorm is building three AI agents — on the OpenClaw framework — to make OpenWorm responsive and welcoming to new contributors, with onboarding, mentoring, and review available 24/7. This page describes the strategy, the agents, and the community model they will enable.
Governing documents
- Contributor Progression Model (Contributing section) — L0–L5 levels, meritocratic ladder, badge system
- DD011: Simulation Stack Architecture — integration support from Mind-of-a-Worm (Design Document, drafted on GitHub)
- AI Contributor Model (Contributing section) — autonomous AI agents as registered contributors
The Problem¶
OpenWorm has operated for over fifteen years as a volunteer-driven, citizen science consortium with 90+ contributors from 16 countries, no central funding, and a single creative leader who holds the scientific vision. This model has produced remarkable results, but it has a structural flaw: new contributors arrive faster than human mentors can onboard them, so enthusiasm decays before it becomes contribution.
As Karl Fogel observes in Producing Open Source Software:
"The price of success is heavy in the open source world. As your software gets more popular, the number of people who show up looking for information increases dramatically, while the number of people able to provide information increases much more slowly."
The degradation is insidious. Experienced contributors silently disengage. Newcomers remain and continue asking questions. The handful of experienced contributors become the only people who can answer them, so they drown in support requests and stop contributing themselves.
Using Nadia Eghbal's project typology from Working in Public (2020), OpenWorm is a "Club" — low user growth, high contributor growth, serving a niche scientific community. Every new member expects personal attention and mentorship from the leadership, because the community feels intimate. That does not scale.
The Three-Ring Model¶
Drawing on the "Megachurch" model described in Open Source Security (January 2026) — a middle ground between Raymond's Cathedral (closed, controlled) and Bazaar (open, chaotic) — we use three concentric rings of participation:
┌─────────────────────────┐
│ Ring 1: Open │
│ Courtyard │
│ │
│ N2-Whisperer answers │
│ questions & orients │
│ newcomers. │
│ │
│ ┌─────────────────┐ │
│ │ Ring 2: │ │
│ │ Proving Ground │ │
│ │ │ │
│ │ Mind-of-a-Worm │ │
│ │ reviews PRs, │ │
│ │ assigns tasks, │ │
│ │ tracks growth │ │
│ │ │ │
│ │ ┌───────────┐ │ │
│ │ │ Ring 3: │ │ │
│ │ │ Inner │ │ │
│ │ │ Sanctum │ │ │
│ │ │ │ │ │
│ │ │ Founder + │ │ │
│ │ │ Mad-Worm- │ │ │
│ │ │ Scientist │ │ │
│ │ └───────────┘ │ │
│ └─────────────────┘ │
└─────────────────────────┘
Ring 1: The Open Courtyard (AI-Gated)¶
Who: Anyone who discovers OpenWorm — students, curious developers, scientists from adjacent fields.
Experience: They interact primarily with the N2-Whisperer AI agent. N2-Whisperer answers questions, explains the project, assigns orientation tasks, and evaluates readiness to move inward.
Access: Read-only access to code and documentation. Can chat in public Slack channels. Cannot open pull requests.
Purpose: Filter and orient. Most visitors get their questions answered and leave satisfied. Those who demonstrate sustained interest and basic competence earn an invitation to Ring 2.
Ring 2: The Proving Ground (AI-Mentored)¶
Who: Contributors who have completed orientation tasks and demonstrated baseline competence.
Experience: They work on real issues under the guidance of the Mind-of-a-Worm AI agent. Mind-of-a-Worm reviews their code against Design Documents, assigns graduated-difficulty tasks, and tracks their progression through competency levels.
Access: Can open pull requests. Can modify documentation. Can contribute to designated "contributor-ready" subsystems. Work is reviewed by Mind-of-a-Worm first, then by a human subsystem maintainer.
Ring 3: The Inner Sanctum (Founder + Trusted Lieutenants)¶
Who: Senior contributors, subsystem maintainers, the Scientific Advisory Board, and the founder.
Experience: L4+ contributors handle architectural decisions; the Mad-Worm-Scientist AI agent aggregates activity from Rings 1 and 2 into a daily summary so the L5 layer doesn't have to monitor every channel.
Access: Full commit access in their subsystem. Can review and merge contributions from Ring 2. Can make architectural decisions within their domain, subject to Design Documents.
The Three AI Agents¶
Each agent is implemented as an OpenClaw skill deployed in the OpenWorm Slack workspace. OpenClaw is an open-source AI agent framework (147K+ GitHub stars) that can be deployed in Slack in approximately 20 minutes.
N2-Whisperer (Newcomer Concierge)¶
Deployment: Public Slack channels (#general, #introductions, #get-started)
| Function | Description |
|---|---|
| FAQ Response | Answers common questions using ingested documentation, papers, and past Slack conversations |
| Project Orientation | Explains the sub-project structure, recommends starting points based on background |
| Skill Assessment | Asks about programming languages, scientific background, and interests |
| Task Assignment | Assigns orientation tasks from a curated list; verifies completion |
| Readiness Evaluation | Evaluates whether newcomers are ready for Ring 2; notifies Mind-of-a-Worm |
| Resource Linking | Points to relevant documentation and past discussions instead of re-explaining |
| Skill Routing | Points you to the right OpenWorm Skill to run on your own machine — N2-Whisperer prescribes the skill; it never writes your code or solves the issue for you |
What it solves: The bottleneck where newcomers wait days (or forever) for an answer to "How do I get started?" and "What should I work on?"
Two meanings of \"skill\" on this page
The three agents above are each built from an OpenClaw agent skill (a SKILL.md that defines the agent's behavior — worm_guide, worm_mentor, worm_digest). Those are different from the OpenWorm Skills that N2-Whisperer routes you to — structured workflows you install into your own AI coding agent and run locally to produce a contribution. See How Contributing Works for the contributor-facing skills.
Mind-of-a-Worm (Active Contributor Guide)¶
Deployment: Contributor-facing channels (#development, #c302, #sibernetic) + GitHub webhooks
| Function | Description |
|---|---|
| PR Pre-Review | Checks contributions against Design Documents, coding standards, and test requirements before a human reviews |
| Design Document Enforcement | Flags deviations from established architectural decisions; links to the relevant DD |
| Graduated Task Assignment | Maintains a 5-level difficulty scale; recommends tasks based on contributor history |
| Progress Tracking | Tracks each contributor's level (L1–L5), completed tasks, review quality |
| Integration Review | Verifies PRs don't break coupling interfaces; tags affected subsystem maintainers (DD011) |
| Peer Mentoring | Connects senior contributors with newcomers working in the same subsystem |
| Badge Issuance & Approval | Awards BadgeList badges — automatically on qualifying GitHub events, and by surfacing manual submissions for prompt review — so contributors who finish the work are never left waiting on a stalled approval queue |
| Badge Health & Upkeep | Monitors each badge's completion funnel (earned vs. in-progress); flags badges with high drop-off or stale/broken instructions and proposes corrected, clearly-scoped, chunked criteria |
What it solves: Code review backlogs that block contributors from seeing their work merged, the gap where new contributors don't know what to work on next, and an unattended recognition layer where finished work goes un-awarded and badge instructions silently rot.
Badge stewardship mandate¶
BadgeList is OpenWorm's recognition layer (live since 2016, 18 public badges, 160+ members), and Mind-of-a-Worm owns its health. Recognition only motivates if it is timely, accurate, and earnable — an unattended badge system actively demotivates. MoaW therefore holds three standing badge duties:
- Surface approvals. Pending badge submissions are reviewed and awarded promptly so a contributor who finishes the work is never stranded waiting. This is a real, observed failure mode: the Simulation Stack Apprentice badge accumulated ~18 completion submissions but only 8 awards during the project's dormant period — roughly half of the people who finished were never credited because no one was there to award them. MoaW closes that gap, and clears any standing backlog of un-awarded finishers.
- Keep instructions current. Badge instructions drift out of date (e.g. Docker setup steps that no longer match the current
openworm/openwormimage or Apple-Silicon realities). When a badge's funnel shows high drop-off, or its steps reference broken/stale tooling, MoaW flags it and proposes corrected, clearly-scoped, chunked criteria with an unambiguous finish line. - Watch the funnel. MoaW tracks earned-vs-in-progress per badge (via the BadgeList API — see the Contributor Progression Model), distinguishes genuine attempts from one-click joins, identifies stalled cohorts, and routes them into re-engagement rather than letting "in-progress" counts accumulate as a graveyard.
Mad-Worm-Scientist (Executive Decision Triage)¶
Deployment: Private channel (#core-digest), visible only to Ring 3 members
| Function | Description |
|---|---|
| Daily Activity Summary | Aggregates all Slack, GitHub, and email activity into a structured digest |
| Decision Queue | Surfaces only items requiring L5 architectural input — design decisions, cross-cutting conflicts, L4 promotions |
| Contributor Radar | Highlights contributors who are rising, fading, or stuck |
| Metrics Dashboard | Reports weekly on conversion rates, PR throughput, and response times |
| Escalation Filter | Routes questions through N2-Whisperer and Mind-of-a-Worm first; only escalates if AI cannot resolve |
What it solves: Avoiding the anti-pattern where the L5 layer has to read every Slack channel to stay informed — which doesn't scale past a handful of contributors.
The Graduated Access System¶
Drawing on the Apache Software Foundation's contributor ladder, the Linux kernel's maintainer tree, and the medical residency model of graduated autonomy, contributors will progress through explicit L0–L5 levels. The canonical table and progression criteria live on the Contributor Progression Model page.
Key principles:
- Merit is earned through sustained contribution, not claimed through enthusiasm
- Levels are subsystem-specific — being an L3 in Sibernetic gives no special status in c302
- Mind-of-a-Worm will track progression automatically and recommend promotions
- Architectural decisions concentrate at L4+ — this is what lets L1–L3 contributors get fast feedback without escalation backlogs
Design Documents as Leverage¶
OpenWorm externalizes its decision-making logic into written Design Documents so contributors don't need to interrupt subject-matter experts for context. These serve the same function as Architecture Decision Records (ADRs) combined with the vision-encoding role of a film director's pre-production materials.
Each Design Document encodes:
- Context — What biological system is being modeled and why
- Decision — The chosen approach, with specificity about fidelity, parameters, and data sources
- Alternatives Considered — What was rejected and why (prevents re-proposals)
- Quality Criteria — How to validate correctness
- Integration Contract — How the subsystem connects to the rest of the organism (DD011)
- References — Relevant papers, datasets, and prior work
Mind-of-a-Worm enforces Design Documents automatically during PR review. Contributors who disagree with a decision can propose a new DD through the RFC process.
Deployment Architecture¶
All three agents run as a single OpenClaw Docker container connected to the OpenWorm Slack workspace:
┌─────────────────────────────────────────┐
│ OpenWorm Slack Workspace │
│ - #general, #introductions │ ← N2-Whisperer
│ - #development, #c302, #sibernetic │ ← Mind-of-a-Worm
│ - #core-digest (private) │ ← Mad-Worm-Scientist
└────────────┬────────────────────────────┘
│ Slack API (WebSocket)
v
┌─────────────────────────────────────────┐
│ OpenClaw Docker Container │
│ ├── skills/worm_guide/ (N2-Whisperer)
│ ├── skills/worm_mentor/ (Mind-of-a-Worm)
│ └── skills/worm_digest/ (Mad-Worm-Scientist)
└────────────┬────────────────────────────┘
│ LLM API
v
┌─────────────────────────────────────────┐
│ LLM Backend (Claude or GPT-4) │
│ Processes SKILL.md instructions │
│ + foundational knowledge docs │
│ + real-time Slack context │
└─────────────────────────────────────────┘
- Infrastructure: DigitalOcean App Platform or self-hosted VPS (~$12–25/month)
- LLM costs: ~$50–150/month at estimated 50–100 API calls/day
- Knowledge base: ~400 pages across 46 files (Design Documents are the bulk)
Precedents¶
This approach is informed by several production deployments:
| Precedent | Key Result |
|---|---|
| Dosu on Apache Superset | Median response time dropped from 2 days to <5 minutes; AI handled 66% of issue triage |
| Gravity Spy (LIGO/Zooniverse) | Personalized difficulty levels improved volunteer accuracy from 54% to 90% |
| Milvus + OpenClaw | Community assistant deployed in 20 minutes; closest existing analog to N2-Whisperer |
| OSSerCopilot (FSE 2025) | 32 design strategies for AI mentors in OSS onboarding |
AI-Assisted vs. AI-Native¶
This page describes the AI-assisted model where AI agents help human contributors. OpenWorm also explores an AI-native model where autonomous AI agents register as independent contributors — see the AI Contributor Model for that complementary approach.
Implementation Status¶
| Component | Status |
|---|---|
| Design Documents (DD002–DD024, DD012.1, DD012.2) | Complete (29 documents, ~350 pages) |
| Agent architecture specification | Complete |
| Agent foundational knowledge docs | Partially complete (~60%) |
| OpenClaw deployment | Not yet deployed |
| Slack integration | Not yet configured |
| GitHub webhook integration | Not yet configured |
The contributor-facing OpenWorm Skills — the workflows N2-Whisperer routes you to — are open source in the openworm/openworm-skills repository. The agents' own behavioral specifications and deployment configurations are maintained privately by the OpenWorm Foundation.
References¶
- Fogel, K. (2024). Producing Open Source Software (2nd edition).
- Eghbal, N. (2020). Working in Public: The Making and Maintenance of Open Source Software. Stripe Press.
- Sarma, G.P. et al. (2018). "OpenWorm: overview and recent advances." Phil. Trans. R. Soc. B 373:20170382. DOI: 10.1098/rstb.2017.0382
- Zevin, M. et al. (2016). "Gravity Spy: Integrating Advanced LIGO Detector Characterization, Machine Learning, and Citizen Science." arXiv:1611.04596. arXiv
- Tan, X. et al. (2025). "OSSerCopilot." ACM/FSE 2025. DOI: 10.1145/3715767