14 Principles. 23 Real Corrections. One Systematic Approach to Training AI Agents.
"Good things are summarized, not invented." — META-00
Your AI agent makes the same mistake twice. You correct it. It says "understood." Next session — same mistake.
You're not alone. AI agents don't learn from corrections by default. Every session is a blank slate. Every fix is temporary.
We solved this.
META is a 14-principle methodology for training AI agents through systematic correction loops. Instead of just fixing errors one at a time, META turns every correction into a permanent behavioral change.
Correction → Pattern Recognition → Abstraction → Injection → Auto-Detection
The result: an agent that genuinely gets better over time. Not claiming to learn — actually learning.
This isn't theory. We logged 23 real corrections to a production AI agent over 2 months. Here's what happened:
| Stage | Description | Automation |
|---|---|---|
| Stage 1 | Manual fixes, no record | 0% |
| Stage 2 | Corrections logged, manually applied | 20% |
| Stage 3 | Patterns auto-detected, Skills auto-updated | 60% |
| Stage 4 | Agent self-corrects in real-time | 90% |
| Stage 5 | Meta-evolution: the evolution itself evolves | 99% |
Our agent moved from Stage 1 → Stage 3 in 8 weeks. 23 corrections. 14 principles. One systematic approach.
| # | What Was Wrong | Root Pattern | META Principle |
|---|---|---|---|
| #1-5 | Repeatedly invented fake e-commerce platform concepts | Hallucination in unfamiliar domains | META-01: 3D Reflection |
| #8 | Misidentified a one-off discount as "channel price chaos" | Single data point → false generalization | META-11: Data Intuition L7 (Sampling) |
| #11 | Claimed "training complete" with zero actual module calls | Confusing reasoning with output | META-10: Quality Control |
| #20 | Used Product B sales data as Product A data | Didn't verify product-line attribution | META-11: Data Intuition L6 (Cross-validation) |
| #21 | Wrote manual config files instead of using the tool's config command | Skipped the simplest path, tested 5 complex ones | META-04: Scalpel Method |
Full corrections log: corrections.md
- META-00 OTF + JIT + Bootstrap: The meta-principle. Summarize on-the-fly, deliver just-in-time, let each iteration produce the next.
- META-01 3D Reflection: Audit by chapter, by type, globally. Iterate until no new findings.
- META-02 Iterative Workflow: P0 Cold Start → P1 Pattern Accumulation → P2 Convergence → P3 Final Verification.
- META-03 Cold Start: Five stages from zero to production. Rules over LLM where possible.
- META-04 Scalpel Method: Decompose tasks. Deterministic rules for what can be ruled. LLM only where necessary.
- META-05 10x Compression: Extract 1 essence from 10x data. Context compression ratio target: 5x.
- META-06 Skill Evolution: Progressive loading → periodic pruning → lifecycle management.
- META-07 Human-First Readability: Humans read first, machines second. Token > XML. Markdown + JSON dual-track.
- META-08 Annotation Engineering: Semantic markup with disambiguation. Closed types (rules) vs open types (LLM).
- META-09 Prompt Engineering: Five-layer pyramid — Goal → Constraints → Method → Verification → Output Spec.
- META-10 Quality Control: Five-layer inspection — Completeness → Grammar → Data → Logic → Cross-process.
- META-11 Data Intuition: Ten layers of data perception. From quality (L1) to decision termination (L10). Most agents fail at L7: sampling raw data before trusting summaries.
- META-12 Data Fusion: Alias detection → disambiguation → bidirectional dedup indexing.
- META-13 Skill Transfer: Three-layer migration — Universal principles → Domain adaptation → Concrete implementation.
- META-14 Structural Granularity: Not all knowledge needs the same structure. From plain text to knowledge graphs is a continuous spectrum.
We don't just claim improvement. We measure it.
META-00 (OTF+JIT+Bootstrap): 17/30 → Stage 2→3
META-07 (Readability): 31/50
META-08 (Annotation): 22/40
META-11 (Data Intuition): 58→56→49→51 (standard tightening)
Raw scores in data/evolution-curve.json.
Create corrections.md. Every time your agent makes a mistake:
What was wrong → What was correct → Root pattern → Permanent rule
Group corrections by pattern. You'll find 3-5 recurring failure modes.
Turn each pattern into a pre-response check. Inject into system prompt, memory, or skill files.
Write auto-detection rules for the most common patterns. When the agent is about to repeat a known mistake — catch it before output.
Track correction frequency. If it's dropping, META is working. If the same correction appears twice, your injection failed.
Correction #20: Agent used Product B data as Product A data.
OTF (On-The-Fly): Logged immediately.
#20 Product B data misattributed to Product A — cross-module quantitative data must verify product-line attribution before integration
Pattern Recognition: "Trusting aggregate labels without sampling raw data" → This is META-11 L7 (Sampling).
Abstraction: Rule: "Before using any quantitative data, sample 3 raw records to verify."
Injection: Added to the agent's pre-response checklist. Now fires before any analysis involving numbers.
Result: Never happened again.
AI agents are powerful but unreliable. The industry focuses on bigger models and better prompts. We focus on something simpler: making agents learn from their mistakes.
This methodology turned an agent that repeated the same 5 errors into one that catches 60% of its own mistakes before output.
No new model required. No expensive fine-tuning. Just systematic correction loops.
meta-agent-training/
├── README.md ← You are here
├── README_CN.md ← 中文版
├── LICENSE ← MIT
├── corrections.md ← 23 real corrections with annotations
├── meta/ ← 14 META principles
│ ├── META-00-otf-jit-bootstrap.md
│ ├── META-01-reflection.md
│ ├── ...
│ └── META-14-structural-granularity.md
├── data/
│ └── evolution-curve.json ← Self-assessment scores over time
├── examples/
│ └── real-corrections.md ← Annotated corrections with META mapping
└── assets/
Have you trained your agent using systematic correction loops? We want your data. Open an issue with your corrections log and we'll add it to the evolution curve.
MIT © 2026
Built from 23 real mistakes. No theory. All practice.