Agent Self-Improvement Harness (7/12) — Heartbeat v2: Multi-Mode State Machine and Escalation

Designing a notification system as a state machine

ํ•ต์‹ฌ ์š”์•ฝ

What this post covers: - Designing Heartbeat as a state machine — decomposing "alive" into a set of modes and state transitions. - Locking escalation to explicit, enumerated reasons — eliminating ambiguous alerts and treating "unknown" as a first-class citizen. - Reducing false-positives with a single parameter change — raising the cold-start timeout from 60s to 120s eliminates model-loading timeouts. - Proactive Preferences feedback loop — feeding user response and non-response into an EMA to dynamically adjust escalation thresholds.

v1 Limitations and v2 Design Goals

The v1 heartbeat operates on a binary state: alive / dead. The structural limitation is that it cannot represent alive but abnormal. This results in an accumulation of alerts where anomalies are detected but root cause is unclear.

v2 has three design goals: 1. Decompose state into the product of state × context 2. Fix alert conditions to explicit enums 3. Incorporate user response as a learning signal

Technique 1: Multi-Mode State Machine

v2 modes follow the {state}.{context} naming convention.

State Cluster Modes
idle idle.normal, idle.degraded, idle.silent
working working.normal, working.slow, working.stuck
recovering recovering.from-crash, recovering.from-quota, recovering.from-network
escalation escalating, escalated, cooling-down
maintenance maintenance.scheduled, maintenance.unplanned
fallback unknown

How It Works

What matters is not the number of modes but the state transition graph. Example: working.normal → working.slow → working.stuck → escalating. Alerts are emitted only at state transition points, not at arbitrary intervals. This constraint structurally prevents repeated alerts for the same persisted state.

Technique 2: Multi-Stage Escalation Reason Enum

Alert conditions are locked to a set of explicit, enumerated reasons:

  1. Quota exceeds 80%
  2. Quota exceeds 100%
  3. Same error repeated 3 times
  4. Response time p95 threshold exceeded
  5. Cron job fails on 2 consecutive runs
  6. Memory directory size runaway
  7. Embedding server unresponsive
  8. External API returning 5xx consecutively
  9. Self-review blocking pattern detected
  10. Retain-tag validation failure rate spike
  11. Zero user response over extended period
  12. New agent self-diagnosis failure
  13. Entry into unknown mode

The Role of Reason 13

Reason 13 is the most critical entry by design. It is the mechanism that promotes indeterminate state to a first-class citizen. A system that stays silent when classification fails carries greater latent risk than a system that fails on classifiable conditions. Designating unknown entry as an escalation reason establishes a path for the system to report "I don't know what I don't know."

Technique 3: Cold-Start Timeout Tuning

The dominant source of false-positive alerts was not complex logic but a single timeout value.

  • Symptom: First heartbeat call times out → classified as no response → escalation
  • Root cause: Model cold-start loading time (initial weight loading + warmup) exceeds the default 60s timeout
  • Fix: Separate cold-start-specific timeout, raised from 60s → 120s
  • Result: False-positives on this path eliminated

Generalizable Pattern

The principle that generalizes from this finding: cold and warm path timeouts must be separated into distinct constants. Handling both paths with a single timeout value means optimizing one path degrades the other. A full analysis of 60 operational failures is covered in Part 16.

Technique 4: Proactive Preferences Feedback Loop

Rather than keeping escalation thresholds static, user response is used as a signal to adjust them dynamically.

Input Signals

  • No response / "quiet" signal: raise the escalation threshold for that mode
  • "Why wasn't I notified" signal: lower the escalation threshold for that mode

Learning Parameters

  • Exponential Moving Average (EMA) + 14-day window
  • Learning rate too fast → alert misses (false-negatives)
  • Learning rate too slow → repeated user correction requests
  • A 14-day window balances convergence and responsiveness

Measured Outcome

EMA-based convergence structurally eliminates repeated threshold-tuning requests compared to static thresholds. However, the first two weeks of the learning period require significant user feedback — that upfront cost is the price for a quieter system afterward.

Limitations and Porting Direction

Current Limitations

  • Potential misbehavior during the initial 2-week learning period
  • High frequency of reason-13 (unknown) entries increases alert fatigue — sub-classification of unknown required
  • EMA window length (14 days) is an empirical constant; re-tuning required per domain

Hermes Port

  • Heartbeat trigger: cron + on_turn_start hook
  • State machine / escalation enum: ported as-is
  • Proactive loop: MemoryProvider.on_memory_write records user response patterns to memory → escalation threshold calculation references this memory

Applicability and Open Questions

Where This Design Applies

  • Systems where alert frequency directly impacts user satisfaction
  • Operational environments where state classification is feasible (observability in place)
  • Channels where user feedback can be collected

Open Questions

  • How far can the unknown mode entry rate be reduced?
  • Is there a better learning curve than EMA (e.g., Kalman filter)?
  • In multi-user environments, how should individual thresholds be separated from shared thresholds?

The essence of a good notification system reduces to: can you explicitly specify the conditions under which you will not alert? The mode set and escalation enum are that specification. Cold-start tuning and the EMA feedback loop are the mechanisms that protect that specification from misfiring.

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 »