Agent Self-Improvement Harness (5/12) — skill-evolve: How an Agent Builds and Fixes Its Skills

Automatic pattern analysis after complex tasks, promotion via repetition threshold, and infinite-loop prevention

Summary

  • After completing a complex task (5+ tools, error recovery), the agent automatically analyzes patterns and registers skill candidates.
  • Only patterns with sufficient repetitions are promoted to skills — a low threshold causes coincidences to become skills.
  • Skills self-improve through a trial-apply → issue-detect → immediate-patch loop.
  • Self-improvement triggers exclusively after real tasks — anything else causes an infinite loop.

Background

Agent skills are typically authored by humans. A developer writes "in this situation, follow this procedure."

The problem is scale. As the variety of tasks an agent handles grows, maintaining every skill manually becomes impractical. Worse, the patterns an agent discovers during actual work — "this approach worked" — are often impossible for a human to anticipate in advance.

The question then becomes: what if an agent could generate its own skills? Detect recurring patterns during complex work, and formalize them into reusable skills.

In OpenClaw, this was implemented under the name skill-evolve.

Body

1. Trigger Conditions — Not Every Task Gets Analyzed

Running pattern analysis after every task is wasteful. A single-line edit or a trivial file change contains nothing worth extracting.

Two trigger conditions were defined:

  • Tasks using 5 or more distinct tools. Multi-tool combinations signal procedure, not simple repetition.
  • Tasks that include error recovery. A sequence that hit an error and resolved it is inherently valuable knowledge.
  • Either condition being met triggers automatic pattern analysis after task completion.

This single filter reduced the analysis target to roughly 20% of all tasks. The remaining 80% were simple tasks that needed no skill extraction.

2. skill-candidates.json — Candidate Registration and Repetition Threshold

A detected pattern does not immediately become a skill. It is first registered as a candidate in skill-candidates.json.

{
  "candidates": [
    {
      "pattern": "SQLite migration with column rename",
      "occurrences": 2,
      "first_seen": "<initial_stage>",
      "last_seen": "<recent_stage>",
      "status": "candidate",
      "tool_sequence": ["read_file", "execute_sql", "write_file", "verify_schema", "run_test"],
      "error_recovery": "column mismatch → ALTER TABLE → re-migrate"
    }
  ]
}

The key design decision is the repetition threshold. The initial value was set low (2 occurrences). The results were poor.

Problem with a low threshold: Coincidences became skills. Two consecutive API errors with similar symptoms were caused by entirely different root causes — one was a timeout, the other an expired auth token. Pattern detection treated them as the same pattern, and the resulting skill proposed the wrong fix.

After raising the threshold: Higher repetition count substantially reduces the probability of coincidence. It signals that the pattern recurs structurally. Incorrect skill generation dropped to near zero.

OpenClaw currently uses 3 occurrences as the baseline. Two was too low; five was too high and suppressed practical skill generation. Three is the empirically derived minimum that qualifies as "sufficiently repeated." Depending on the project's task volume, four may be appropriate.

3. Trial Application and Immediate Patching — Skills Have Bugs Too

Once a candidate crosses the threshold and is promoted, it does not go straight to production. It enters a trial application phase.

The next time a matching task arrives, the new skill is applied. The result is then verified against three criteria:

  • Did the procedure the skill proposed actually succeed?
  • Were there any unexpected side effects?
  • Were any steps missing?

If a problem is found, the skill file is patched immediately and the reason for the change is recorded. This is what "self-improvement" means in practice. A skill is not permanent once created — it is validated and refined with every application.

Real example: when a file migration skill was first created, the "create backup" step was absent. During trial application, a migration failure left no way to recover the original files. "Step 0: Backup source" was added immediately.

4. Infinite Loop Prevention — The Self-Improvement Trap

Self-improvement introduces a fundamental risk: a loop that improves the improvement process itself.

If pattern analysis is itself detected as a complex task, the system starts analyzing the analysis, then analyzing that analysis — an unbounded loop.

The prevention rule is simple: self-improvement triggers only after a real task.

Real task (file edit, code generation, bug fix, etc.)
  → Pattern analysis ✓ (triggered)

Pattern analysis itself
  → Pattern analysis ✗ (not triggered)

Skill patching itself
  → Pattern analysis ✗ (not triggered)

"Real task" is defined precisely: a task initiated by a user request that produces changes to project files. Meta-tasks — skill management, memory cleanup, pattern analysis — are excluded from the trigger scope.

Lessons Learned

Skill explosion. When the threshold was low, skill candidates proliferated rapidly. Most were useless, and managing the candidate list itself became a cost. Raising the threshold cut candidate generation rate by more than half and raised average quality.

Over-generalization. "Read a file, modify it, save it" matches nearly every task. Skills built on overly general patterns have no value. The minimum tool sequence length (5 steps) and the error-recovery requirement filtered out this class of candidates.

Token runaway during Hermes migration. When migrating self-improvement logic from OpenClaw to Hermes, the trigger rules were not ported first. The self-improvement loop began firing on meta-tasks, and token usage exploded. OpenClaw was restored and the trigger isolation rules were hardened. A Hermes retry is now under validation. When migrating a self-improvement architecture, port the trigger isolation rules before anything else.

Conclusion

The skill self-improvement loop gives an agent the ability to learn from experience.

  1. Restrict the trigger. Analyze only complex tasks — 5+ tools used or error recovery present.
  2. Promote after sufficient repetition. The minimum criterion that distinguishes structure from coincidence. OpenClaw's current baseline is 3.
  3. Trial-apply, then patch. Skills have bugs like code does. Fix them in use.
  4. Meta-tasks do not trigger. Infinite self-improvement loops are blocked at the source.

Given enough operating time, skills the agent generated from real work become more practical than the ones a human authored in advance — because the agent encountered those exact situations firsthand.

Series overview: Series index

๋Œ“๊ธ€

๋Œ“๊ธ€ ์“ฐ๊ธฐ

์ด ๋ธ”๋กœ๊ทธ์˜ ์ธ๊ธฐ ๊ฒŒ์‹œ๋ฌผ

Agent Memory Engine (2/10) — Building an AI Agent Memory System with SQLite Alone

← 1/10 SQLite Memory Engine f… ๐Ÿ“š Series Index 3/10 memcore Completed → A memory engine that runs without daemons, without dependencies, anywhere ํ•ต์‹ฌ ์š”์•ฝ This post covers how to design an AI agent memory engine (memcore) on a single-file SQLite backend. - Structure: 25 modules, ~6,464 lines, 22 DB tables - Principles: no daemon, single-file portability, no required external dependencies, host-agnostic - Core techniques: 4-layer hybrid search (topic · vector · FTS5 · LIKE), U-tag dialectic-based confidence evolution, ingestion-layer bias rejection rules What This Post Covers One approach to building AI agent memory without a server or cloud vector DB. Specifically: (1) the limitations of text-file-based memory, (2) vector DB alternatives compared and selection criteria, (3) how to layer search tiers, (4) how to evolve confidence scores over time, and (5) how to block biased memory accumulation at the ingestion stage. Design Principles: Why Single-File SQLite Text-file-based ...
Read more »

"ML Foundations (9/9) — PyTorch vs TensorFlow, and the Road to Local LLMs"

์ด๋ฏธ์ง€
← 8/9 Deep Learning Architec… ๐Ÿ“š Series Index (series end) The final part. So far we've focused on models. Now we focus on the tools that actually run them . We'll lay out the philosophical difference between PyTorch and TensorFlow, and trace how Hugging Face transformers , llama.cpp , MLX , and Ollama built the bridge to running large language models on your own machine. By the end you should have the full mental model of "download a pretrained LLM and serve it locally." 0. Learning Objectives Explain the eager-vs-graph difference between PyTorch and TensorFlow. Explain, in graph terms, how autograd automates the backward pass. Use Hugging Face transformers ' AutoModel , AutoTokenizer , and pipeline abstractions. Describe llama.cpp's GGUF quantization, INT4 inference, and CPU-first flow. Describe Apple MLX's use of unified memory and how it differs from PyTorch. Run a local LLM with Ollama and call it through an OpenAI-compatible API. ...
Read more »

"RAG Core Study (14/26) — Evaluation Sets with RAGAS & DeepEval"

์ด๋ฏธ์ง€
Series overview: Series index ๐Ÿ“š Series Index Without an evaluation set, every RAG improvement is a story, not evidence. RAG systems fail in more than one place: retrieval, context selection, answer generation, and grounding. That means "looks better" is not a metric. Part 14 explains how to build a golden dataset , how RAGAS and DeepEval fit into the workflow, and why teams must separate retrieval quality from answer quality if they want tuning decisions to hold up. 0. Prerequisites Part 13 reranking — system quality now depends on multi-stage ranking. Part 1 grounding — a correct-looking answer is not enough. Part 12 Hybrid — multiple retrieval settings need comparison on fixed data. 1. Learning Objectives Build a small but useful golden dataset for RAG. Distinguish retrieval metrics from answer metrics . Use RAGAS and DeepEval in the right roles. Avoid the common traps in judge-model-based evaluation. 2. ํ•ต์‹ฌ ์š”์•ฝ An evaluation set for RAG is ...
Read more »

"ML Foundations (8/9) — Deep Learning Architectures: CNN, RNN, Attention"

์ด๋ฏธ์ง€
← 7/9 Deep Learning Training ๐Ÿ“š Series Index 9/9 PyTorch vs TensorFlow → What did we keep adding on top of an MLP? Vision went CNN, sequences went RNN/LSTM, and eventually both converged on Attention and the Transformer. Each architecture is easier to remember if you read it as what it gave up to gain something else . This part is the one-line summary of the decisive inflection points: LeNet → AlexNet → ResNet → LSTM → Attention → Transformer. 0. Learning Objectives Explain why CNN's convolution, pooling, and weight sharing match images so well. Trace what got unblocked from LeNet → AlexNet → VGG → ResNet as depth grew. State BPTT for RNNs and the vanishing/exploding gradient problem. Write the LSTM gate equations (forget, input, output) with intuition. Write the attention formula in query/key/value form and the scaled dot-product variant. State the precise reasons Transformers replaced RNNs (parallelism + long-range dependencies). 1. ํ•ต์‹ฌ ์š”์•ฝ CNN : weight sharin...
Read more »

"ML Foundations (7/9) — Deep Learning Training: Optimizers, Regularization, Initialization"

← 6/9 Neural Networks ๐Ÿ“š Series Index 8/9 Deep Learning Architec… → Building an MLP, as we did in Part 6, is not the same as getting it to train well . The right optimizer, regularization, initialization, and learning rate — when these line up, deep networks converge. When they don't, the network refuses to learn at all. This part is the catalog of those crafts, with formulas, paper citations, and code in one place. 0. Learning Objectives Compare and write the update rules for SGD, Momentum, Nesterov, and Adam. Explain how Dropout, BatchNorm, and LayerNorm work and where they belong in a model. Derive the variance formulas for Xavier and He initialization and match them to activations. Implement step, cosine, and warmup learning-rate schedules in PyTorch. Explain why gradient clipping is effectively required for RNNs and Transformers. Diagnose the most common training failures (NaN, plateau, overfitting) and apply first-line fixes. 1. ํ•ต์‹ฌ ์š”์•ฝ SGD : \(w \leftarro...
Read more »

OpenClaw to Hermes Migration (2/13) — What to Preserve, Partially Port, or Discard

← 1/13 Current Structure Inve… ๐Ÿ“š Series Index 3/13 What Moves to Hermes → A code review record classifying operational scripts into three tiers: preserve, partial-preserve, and discard. What This Post Covers An asset classification framework (preserve / partial-preserve / discard) for migrating a legacy agent system to a successor A concrete case applying the Strangler Fig + Subset Migration strategy to a memory pipeline Classification results evaluated across LOC, external dependencies, and replaceability axes A method for slimming down a bloated single file (1,310 LOC) using the migration as a natural entry point Problem Definition: Port Everything or Cut Deep? The OpenClaw inventory comprises multiple agents, multiple scheduler/daemon processes, and multiple operational scripts. When moving to the successor system, Hermes, the choice is not binary. A full port carries accumulated technical debt along with it; a minimal port discards hard-won domain logic. Code-revie...
Read more »

AI Agents I Built (5/7) — Building an Automated Blogger API Publishing System

← 4/7 My Life Organizer Agent ๐Ÿ“š Series Index 6/7 Dual → Designing a Markdown-Based Auto-Publish Pipeline with Blogger API v3 + OAuth 2.0 Key Summary Blogger API v3 + OAuth 2.0 enables a pipeline that converts Markdown files into publishable HTML posts blogger_publish.py handles the full flow — frontmatter parsing → body cleanup → HTML conversion → CSS injection → API post — in a single script For bulk publishing, quota management (delay, retry, TOC suppression) is the critical design concern Platform Selection Criteria API access is mandatory for AI agent-driven content publishing. A comparison of platforms: Platform API Cost AdSense Decision WordPress.com REST API Paid plan required Paid only Cost overhead Ghost REST API Self-hosted or paid Manual setup Ops overhead Blogger v3 API Free Native integration Adopted Blogger is free, exposes a REST API, and integrates natively with AdSense. For personal blog automation, this combination i...
Read more »