Recovery memory layer for AI tool failures. When your agent's third-party tool path fails, AgentRX surfaces what worked (and what didn't) from prior experience.
AgentRX is a recovery experience library for AI agents. It answers one question:
The agent's tool failed — what did someone else try that worked?
It focuses on:
- Third-party tool failures — playwright, browser-cdp, web-fetch, MCP connectors, etc.
- Environment long-tail issues — proxy, headless, sandbox, permissions, dependencies
- What worked / what didn't — concrete, actionable recovery guidance
AgentRX does not compete with platform-level recovery (retry, circuit breaker, fallback routing). Those are the platform's job. AgentRX handles what platforms can't cover: real-world tool-specific experience.
1. Agent's third-party tool fails
2. Agent tries to recover on its own — fails 2+ times
3. Agent activates AgentRX skill
4. Agent describes the failure in natural language
5. Agent finds the matching tool file in cases/
6. Agent reads the closest case, focuses on what_worked and what_didnt_work
7. Agent applies the learned recovery
8. Agent appends the outcome to ledger.jsonl
The v2.1 architecture (schema validation, route registry, deterministic retrieval) was a well-intentioned infrastructure investment — but it put cart before horse. Before building a retrieval engine, we need to verify: does case content actually help agents recover?
This MVP tests that question with minimal infrastructure. If the answer is "yes", we'll grow back the heavy architecture on a proven foundation. If "no", we saved months of wasted engineering.
- A very short SKILL.md that tells the agent when and how to use AgentRX
- Flat
cases/directory — one JSON file per tool, containing case arrays ledger.jsonlfor append-only experience logging- Minimal case format: symptom, what_worked, what_didnt_work, times_confirmed
- v2.1 JSON schema validation
- Route registry / route recommendation
- Deterministic retrieval scripts
- Build index / validate / generate scripts as main path
- Automated hooks triggering
- Complex directory分层 (by_tool, by_tag, verified, curated, archived)
- Synthetic seed cases
| File | Role |
|---|---|
| SKILL.md | Ultra-short runtime protocol the agent reads when activated |
| cases/ | Flat case directory — one JSON file per tool |
| ledger.jsonl | Append-only experience log |
| CONTRIBUTING.md | How to contribute real recovery cases |
| docs/MVP.md | Why we shrank, what we validate, future expansion order |
| docs/legacy/ | v1→v2.1 architecture exploration (historical reference only) |
MIT