AgentRX

🩺 A task-first stuck-state protocol + deterministic retrieval layer for AI agents. When your agent's tool path fails, AgentRX structures the failure, retrieves similar cases, and surfaces candidate routes.

What is AgentRX?

AgentRX is not a human-facing tool directory. It is a machine-consumable protocol that answers one question:

The agent is stuck — what should it do next?

It provides:

• Schema — a standard v2.1 case format (evidence + inference separation)
• Route registry — stable action paths, not tool brand names
• Validation — JSON Schema + cross-file rule consistency checks
• Indexing — lightweight case library index for retrieval
• Deterministic retrieval — retrieve_cases.py finds top-k candidate cases

AgentRX does not run inference for the agent. Route recommendation is the agent's own reasoning, based on retrieved cases + the route registry.

A concrete example

User: Extract the product list from this page.

AI: [tries browser-cdp skill]
    The page uses JavaScript to render content. browser-cdp only
    returned the initial HTML shell. Data missing.

[AgentRX activates]

AgentRX: Retrieved similar cases → route: switch_to_alternative_tool_path

         Why: current tool captures static HTML only; page requires
         JavaScript rendering.

         Candidate: playwright-mcp can render the page and extract
         the full DOM. web_fetch is a lighter option for static pages.

---

What AgentRX provides today

Component	Status
Case schema (v2.1)	✅
Route registry	✅
Case validation	✅ (JSON Schema + cross-file rules)
Index building	✅
Deterministic retrieval	✅ (retrieve_cases.py)
Case ID generation	✅ (new_case_id.py)

What AgentRX does not provide (yet)

Component	Status
Automated case review / merge / publish pipeline	🚧 planned
Python-based route recommender	❌ out of scope — agent does its own route inference

---

Install

Claude Code

git clone https://github.com/LpcPaul/AgentRX.git ~/.claude/skills/agentrx

OpenClaw / ClawHub

git clone https://github.com/LpcPaul/AgentRX.git ~/.openclaw/skills/agentrx

Codex / Cursor / other skill-compatible runtimes

git clone https://github.com/LpcPaul/AgentRX.git ~/.codex/skills/agentrx

---

How it works

1. AI gets stuck (concrete failure signal)
2. AI structures the stuck state (evidence + inference)
3. AI retrieves similar cases via retrieve_cases.py
4. AI chooses a route based on retrieved cases + rules/routes.yaml
5. AI records the outcome
6. The new case becomes available for future AI agents

Human installs. AI operates.

	What they do
Human	Install the skill. Host the repository. Maintain schema/taxonomy.
AI	Detect stuck state. Collect evidence. Retrieve similar cases. Choose a route. Optionally contribute a new case.

AI contributors must submit complete v2.1 JSON. Human fallback / form-to-JSON assembly is no longer supported.

---

Read this next

Document	Role
SKILL.md	The runtime protocol the AI agent reads when activated
docs/ARCHITECTURE.md	System design — why evidence/inference, why route ids
docs/INTAKE_CARD.md	The structured intake card format
CONTRIBUTING.md	How cases enter the system — JSON-only contribution path
cases/README.md	Case library structure and indexing

License

MIT