Agent Operations Retrospective (1/7) — 7 Lessons Distilled from 60 Operational Failures

cold-start 60→120s, page_gap 8→50, exit-code split — the most powerful fixes were single-line changes

ํ•ต์‹ฌ ์š”์•ฝ

  • Approximately 60 failure logs accumulated during long-term operation were analyzed and distilled into 7 reusable lessons.
  • The most effective fixes were not large-scale refactors — they were single-line threshold or definition changes.
  • The common principle across all 7 lessons: make the system express more precisely what it knows and what it does not know.

This post targets engineers working with failure logs from agent-based systems. It covers 7 parameters, conventions, and cache policies where measurable improvements were observed.

1. Exit-Code Convention Split

The conventional meaning of exit 0 is "everything is OK," but in production, it must be defined as "my job is done" for cron analysis to function correctly. Split into three modes:

  • --strict mode: 0 = validation passed, 1 = validation failed.
  • Reflect mode: 0 = clean exit regardless of queue state.
  • Daily-audit mode: 1 = "issues found" (success), 0 = no results (suspicious).

How it works. The inverted code convention in daily-audit is designed to flag a "check that found nothing" as a suspect signal. It surfaces cases where the inspection script exits silently empty-handed due to sampling failure, missing input, or similar silent faults.

Impact. This single convention change allowed the cron analyzer to distinguish genuine failures from normal exits. Previously, all exit 0 values were treated identically, leaving suspect cases undetected.

2. Automatic Warning Resolution

Warnings that do not recur within 10 days are automatically resolved. The underlying assumption is simple: real problems come back.

How it works. One-off noise — transient network blips, external API rate-limit spikes — does not recur, so it should not remain in the long-term tracking queue. The 10-day window covers at least one retry cycle for periodic cron and batch jobs.

Limitation. Intermittent failures with a recurrence interval exceeding 10 days (e.g., monthly) may be missed by auto-resolution. A separate monthly accumulation counter compensates for this gap.

3. page_gap 8 → 50

A chunk-linking parameter in the memory/ directory. Even when topics match, chunks beyond the threshold distance are not merged into the same graph.

How it works. page_gap=8 requires a topic to appear continuously within 8 chunks — a strict setting. In long sessions where topics resurface intermittently, the graph fragments into disconnected subgraphs. Raising to 50 allows the same topic scattered across a week's worth of sessions to be consolidated into a single graph.

Impact. Recall hit-rate improved noticeably. The trade-off is increased false-positive connections, which are filtered at the topic-matching layer (Resolution L2/L1/L0, §5).

4. Heartbeat cold-start 60s → 120s

The first-call timeout ceiling. Covered in Part 8 but worth reiterating.

How it works. Most cold-start timeouts originate from model loading time — not network latency or queue saturation. The dominant cause is the one-time memory load on first invocation.

Impact. A single line change from 60s to 120s eliminated false-positive alerts during overnight hours. The count of heartbeats misclassified as failures dropped to near zero.

5. Topic-Cued Recall TTL 60min + Resolution L2/L1/L0

Cache policy for retrieval results. A simple 60-minute TTL cache is combined with a resolution-based update mechanism.

Resolution tiers. - L2 = exact match - L1 = topic match - L0 = keyword match

How it works. If the same topic is recalled within 60 minutes, the cache is used. However, if the cached hit has low resolution (L0/L1), a 0.3x score penalty is applied, preserving the opportunity for a higher-precision result to replace it. This structure converts a simple TTL cache into an adaptive cache.

Applicability. Generalizes to any agent system with a retrieval layer. Adjust the TTL value to match the domain's topic-switching frequency.

6. softThresholdTokens 6000 — Early Compression

Context compression trigger. Fire it well before the limit, not near it.

How it works. Triggering compression at the limit means too much content is compressed at once, increasing information loss. Triggering at the 6000-token threshold reduces the volume compressed in each pass, which reduces semantic loss per compression event.

Impact. Frequency of compression degrading model response quality decreased. Because early compression runs in smaller increments, cumulative loss per session follows a gentle curve rather than a linear ramp.

7. Retain 4 Categories + c=X Confidence

Two metadata axes are attached to stored entries: a category tag and a confidence score.

Category tags. - W = world fact - B = behavior pattern - O = opinion - S = source pointer

Confidence c=. A value between 0 and 1. Combined with confidence-decay, it decreases automatically over time.

How it works. The decay function operates differently per category — old opinions (O) weaken automatically; old facts (W) have a decay coefficient near zero and remain nearly stable. This distinction enforces at the system level that facts are stable and interpretations are subject to review.

Common Pattern: Making Room for Not-Knowing

The single principle threading through all 7 lessons:

Give the system a designated place to express that it does not know.

  • exit-code split expresses: "is this a failure or a normal exit?"
  • c= confidence expresses: "how certain is this information?"
  • Resolution L0/L1/L2 expresses: "how precise was this retrieval?"

A system with dedicated representation for uncertainty becomes capable of self-inspection. An interface that does not hide uncertainty is the practical starting point of observability.

Portability: Hermes Port Mapping

All 7 lessons are backend-agnostic. Mapping to Hermes equivalents:

Lesson Hermes Target
exit-code convention cron payload
page_gap memcore parameter
cold-start 120s Hermes hook
TTL 60min + Resolution prefetch cache
softThreshold 6000 context_compressor config
Retain 4 categories ingest pipeline
Warning auto-resolution audit loop

All are structurally portable with a 1:1 mapping.

Summary: Large Refactor vs. Single-Line Fix

Large refactors typically introduce larger bugs. The most effective improvements during long-term operation were single-line threshold or definition changes. Three properties they share:

  1. Locality — the change scope is within one file and one line.
  2. Semantic-layer change — not merely changing a value, but redefining what that value means.
  3. Observability gain — logs carry more information after the change than before.

Open Questions

  • Are 4 category tags (W/B/O/S) sufficient? Long-term operation may surface domains requiring finer granularity.
  • page_gap=50 is optimal at current operating scale, but whether the same value holds for domains with different chunk density (e.g., high-frequency conversation vs. low-frequency documents) requires re-measurement.
  • The decay curve for c= confidence is currently exponential. Non-linear decay (e.g., step function) may be more accurate for certain categories.

Values in this post should be treated as starting points. Adjust to the domain characteristics of the target system.

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 »