# Nick McCarty

> Technical blog covering agentic systems engineering, local LLM inference on consumer hardware,
> self-improving prompt optimization loops, multi-agent orchestration, computer vision,
> and ML research. All harness work runs on a local RTX 5000 Ada with llama.cpp / vLLM backends.

## Main pages

- [Blog index](https://nickmccarty.me/blog.html): Full post listing with category filter and full-text search
- [About](https://nickmccarty.me/about.html): Background, current work, and contact

## Key posts — Agentic Harness Engineering

- [The Harness Thesis: Why Scaffolding Beats Model Selection](https://nickmccarty.me/blog/harness-thesis.html): The central claim of agentic harness engineering: for a fixed task domain, the quality of the scaffolding surrounding a language model matters more than the choice of model
- [The Wiggum Loop: Cross-Model Evaluation and Dimensional Revision](https://nickmccarty.me/blog/the-wiggum-loop.html): The producer-evaluator separation pattern: how two loops—Ralph (outer) and Wiggum (inner)—compose to produce reliably high-quality agentic output
- [Inside agent.py: The Three-Turn Research Pipeline](https://nickmccarty.me/blog/agent-pipeline.html): agent
- [inference.py: The Unified LLM Backend Shim](https://nickmccarty.me/blog/inference-shim.html): inference
- [The Memory View: 2,173 Observations, Quality Signals, and an Ontology Graph](https://nickmccarty.me/blog/memory-view.html): A tour of the harness Memory UI: a semantic observation store that accumulates every run as a structured fact, filterable by quality signal, searchable by task, and visualized as a force-directed ontology graph
- [The Autoresearch View: 40 Experiments, One Keep, and a 3% Signal Rate](https://nickmccarty.me/blog/autoresearch-view.html): A tour of the harness Autoresearch UI: real-time supervision of the SYNTH_INSTRUCTION optimizer
- [The Regression Harness: Eval Suite, Criterion Functions, and Experimental Infrastructure](https://nickmccarty.me/blog/eval-suite.html): How the harness's three-layer measurement stack works: the task registry and criterion function library in eval_suite
- [The Lit-Review Skill: Seven Steps from ArXiv Query to Rendered Survey](https://nickmccarty.me/blog/lit-review-skill.html): lit_review_skill
- [The OSINT Skill: 11-Layer Target Enrichment for Research Tasks](https://nickmccarty.me/blog/osint-skill.html): A taxonomy of the harness OSINT skill: target detection by regex, 11 parallel enrichment layers (DNS, RDAP, crt
- [Agentic System Design Patterns](https://nickmccarty.me/blog/agentic-design-patterns.html): A catalog of 27 named design patterns for agentic LLM systems, organized into eight categories: substrate, context, verification, orchestration, security, observability, and self-improvement
- [The Wiggum Panel: Three-Persona Parallel Evaluation](https://nickmccarty.me/blog/panel.html): WIGGUM_PANEL=1 runs Domain Practitioner, Critical Reviewer, and Informed Newcomer in parallel threads and merges their issue lists into the Wiggum revision loop before the first rewrite
- [Closing the SkillOpt Gaps: What Actually Shipped](https://nickmccarty.me/blog/skillopt-shipped.html): A previous post identified three gaps between the harness autoresearch loop and SkillOpt: proactive validation gating, fast/slow epoch structure, and a persistent skill artifact
- [Building the Detectors: What Actually Shipped After 107 Experiments](https://nickmccarty.me/blog/convergence-detectors-shipped.html): The four convergence detectors proposed after 90 failed experiments were implemented — but not as designed
- [From Hill-Climbing to Pareto: GEPA and the Autoresearch Loop](https://nickmccarty.me/blog/gepa-autoresearch.html): DSPy's GEPA optimizer formalizes exactly what autoresearch
- [Security Patterns: Constraining the Agent](https://nickmccarty.me/blog/security-patterns.html): Four patterns that constrain what the agent can do to the host system and what external content can do to the agent's memory: AST Guard, Path Sandbox, Injection Scanner, and CDP Guard
- [Context Engineering: What Reaches the Model](https://nickmccarty.me/blog/context-engineering.html): Five patterns governing what information reaches the synthesis model — the Planner-First, Novelty Gate, Dual-Backend Memory Store, Semantic Chunker, and Vision Bridge
- [harness-data-model](https://nickmccarty.me/blog/harness-data-model.html): Five JSONL files, one entity hierarchy, and everything tokens_by_stage is actually measuring
- [The Economic Data Tools: FRED and BEA](https://nickmccarty.me/blog/fred-bea-tools.html): fred_tool
- [Harness vs. Perplexity: Four Iterations to a Tie](https://nickmccarty.me/blog/perplexity-vs-harness.html): Eight harness iterations against a frozen Perplexity output on a current-conditions Fed district inflation task
- [The Alt-Data Pipeline: From Beige Book to Paper Trading Thesis](https://nickmccarty.me/blog/alt-data-pipeline.html): Five new enrichment layers — BEA, market signals, yfinance, Alpaca portfolio context, and a structured thesis synthesis template — close the loop from macro narrative to actionable paper trading theses

## All posts by section

### Agentic Harness Engineering

- [The Case Against Prompt Engineering](https://nickmccarty.me/blog/case-against-prompt-engineering.html)
- [Cost Envelope Management for Production AI Agents](https://nickmccarty.me/blog/cost-envelope-management.html)
- [The Economic Data Tools: FRED and BEA](https://nickmccarty.me/blog/fred-bea-tools.html)
- [The Wiggum Panel: Three-Persona Parallel Evaluation](https://nickmccarty.me/blog/panel.html)
- [Closing the SkillOpt Gaps: What Actually Shipped](https://nickmccarty.me/blog/skillopt-shipped.html)
- [The Plugin System: /forge:plugin and /forge:list](https://nickmccarty.me/blog/forge-plugin-skills.html)
- [The Interview Skills: /grill-me and /onboarding](https://nickmccarty.me/blog/onboarding-grill-me-skills.html)
- [The Developer Utility Skills: /scratchpad and /test-harness](https://nickmccarty.me/blog/scratchpad-test-harness-skills.html)
- [The Trading Skills: /validate-trades and /execute-trades](https://nickmccarty.me/blog/trade-execution-skills.html)
- [The Diagnostic Skills: /debug and /troubleshoot](https://nickmccarty.me/blog/debug-troubleshoot-skills.html)
- [The Site Generation Skills: /design and /build-page](https://nickmccarty.me/blog/design-build-page-skills.html)
- [The Site Generation Skills: /site and /deck](https://nickmccarty.me/blog/site-deck-skills.html)
- [The Navigation Skills: /suggest and /re-orient](https://nickmccarty.me/blog/suggest-reorient-skills.html)
- [The Explorer View: Per-Run Pipeline DAG Inspector](https://nickmccarty.me/blog/explorer-view.html)
- [The Page Feedback Widget: Closing the Loop Between Browser and Agent](https://nickmccarty.me/blog/page-feedback-widget.html)
- [The Pipeline View: Data Enrichment DAG](https://nickmccarty.me/blog/pipeline-view.html)
- [The Research History View: A Unified Activity Log](https://nickmccarty.me/blog/research-history-view.html)
- [inference.py: The Unified LLM Backend Shim](https://nickmccarty.me/blog/inference-shim.html)
- [harness/api: The FastAPI Backend and Dashboard Server](https://nickmccarty.me/blog/server.html)
- [The Skills Registry: Hook Points, Auto-Activation, and 38 Skills](https://nickmccarty.me/blog/skills-registry.html)
- [Inside agent.py: The Three-Turn Research Pipeline](https://nickmccarty.me/blog/agent-pipeline.html)
- [The Position Swap: Beige Book RAG Results and the DPO Cold-Start Problem](https://nickmccarty.me/blog/beige-book-rag-dpo.html)
- [Building the Detectors: What Actually Shipped After 107 Experiments](https://nickmccarty.me/blog/convergence-detectors-shipped.html)
- [Live Data Beats Narrative: FRED RAG Experiment Results](https://nickmccarty.me/blog/fred-rag-results.html)
- [Memory as Infrastructure: Quality-Weighted Retrieval and the Ontology Graph](https://nickmccarty.me/blog/memory-observability.html)
- [OSINT Enrichment: Nine Layers of Passive Reconnaissance](https://nickmccarty.me/blog/osint-enrichment.html)
- [The Planner: Two-Pass Pre-Research Analysis](https://nickmccarty.me/blog/planner.html)
- [The Search Cache: SQLite TTL Caching for DDGS Queries and Research Contexts](https://nickmccarty.me/blog/search-cache.html)
- [Deploying the Harness with Docker: CPU, GPU, and Compose Variants](https://nickmccarty.me/blog/docker-deployment.html)
- [The GitHub Skill: LLM-Assisted Git Operations from the Agent Loop](https://nickmccarty.me/blog/github-skill.html)
- [The Subagent Demo Suite: Orchestrating Multi-Task Research Portfolios](https://nickmccarty.me/blog/subagent-demo.html)
- [Synthetic Eval Task Generation with TinyTroupe Personas](https://nickmccarty.me/blog/tinytroupe-tasks.html)
- [The Email Skill: Personalized Outreach Drafts from Conference Speaker CSVs](https://nickmccarty.me/blog/email-skill.html)
- [The Lit-Review Skill: Seven Steps from ArXiv Query to Rendered Survey](https://nickmccarty.me/blog/lit-review-skill.html)
- [Mining a Ground-Truth Knowledge Base for the Eval Suite](https://nickmccarty.me/blog/mine-knowledge.html)
- [The Wiki Sync Skill: Deterministic Source Extraction and Gap-Targeted Code Injection](https://nickmccarty.me/blog/wiki-sync.html)
- [Agentic System Design Patterns](https://nickmccarty.me/blog/agentic-design-patterns.html)
- [Agentic Threat Hardening: The OWASP Top 10, Applied](https://nickmccarty.me/blog/agentic-threat-hardening.html)
- [When the Loop Defeats Itself: Convergence Failures in Autonomous Prompt Optimization](https://nickmccarty.me/blog/autoresearch-convergence.html)
- [The Regression Harness: Eval Suite, Criterion Functions, and Experimental Infrastructure](https://nickmccarty.me/blog/eval-suite.html)
- [The Playwright Skill: LLM-Guided Navigation via ARIA Snapshots](https://nickmccarty.me/blog/playwright-skill.html)
- [SBOM and AIBOM for Agentic Systems](https://nickmccarty.me/blog/sbom-aibom.html)
- [What SkillOpt Gets Right](https://nickmccarty.me/blog/skillopt-gaps.html)
- [YouTube and Media Transcription: Two Paths, One Research Input](https://nickmccarty.me/blog/youtube-transcribe.html)
- [Five Personas, One Veto: Consensus Filtering for Fine-Tuning Data](https://nickmccarty.me/blog/curator.html)
- [The op CLI: A Rich REPL for the Research Harness](https://nickmccarty.me/blog/op-cli.html)
- [The Supervisor: Four Convergence Signals and Advisory Interventions](https://nickmccarty.me/blog/supervisor.html)
- [The Audio Data Flywheel](https://nickmccarty.me/blog/audio-data-flywheel.html)
- [Experimental Methodology: Four Experiments, One Pipeline](https://nickmccarty.me/blog/experimental-methodology.html)
- [From Hill-Climbing to Pareto: GEPA and the Autoresearch Loop](https://nickmccarty.me/blog/gepa-autoresearch.html)
- [The MCP View: Exposing the Harness as a Tool Server](https://nickmccarty.me/blog/mcp-view.html)
- [Multi-Objective Alignment: Beyond Scalar Rewards](https://nickmccarty.me/blog/multi-objective-alignment.html)
- [The Security View: 36 Events, 28 Blocks, and Live Injection Intercepts](https://nickmccarty.me/blog/security-view.html)
- [The Telemetry Router: Mining runs.jsonl for Self-Directed Improvement](https://nickmccarty.me/blog/telemetry-router.html)
- [The Voice View: Push-to-Talk Notes, Task Dictation, and the Audio Data Flywheel](https://nickmccarty.me/blog/voice-view.html)
- [Context Engineering: What Reaches the Model](https://nickmccarty.me/blog/context-engineering.html)
- [A Failure Taxonomy for Agentic Systems](https://nickmccarty.me/blog/failure-taxonomy.html)
- [The Harness Thesis: Why Scaffolding Beats Model Selection](https://nickmccarty.me/blog/harness-thesis.html)
- [Inference Patterns: The Substrate Layer](https://nickmccarty.me/blog/inference-patterns.html)
- [Observability and the Data Flywheel](https://nickmccarty.me/blog/observability-flywheel.html)
- [Orchestration Patterns: Scaling to Multi-Agent Execution](https://nickmccarty.me/blog/orchestration-patterns.html)
- [Parallel Inference: Hardware Substrates for LLM Workloads](https://nickmccarty.me/blog/parallel-inference.html)
- [The Pipeline in Motion: Tracing a Task Through All Eleven Subsystems](https://nickmccarty.me/blog/pipeline-in-motion.html)
- [Security Patterns: Constraining the Agent](https://nickmccarty.me/blog/security-patterns.html)
- [Sessions and Artifacts: The Harness Output Registry](https://nickmccarty.me/blog/sessions-artifacts-view.html)
- [The Wiggum Loop: Cross-Model Evaluation and Dimensional Revision](https://nickmccarty.me/blog/the-wiggum-loop.html)
- [Verification Patterns: Measuring and Improving Output Quality](https://nickmccarty.me/blog/verification-patterns.html)
- [The Fine-tune View: DPO Training Runs, RL Data, and the Preference Feedback Loop](https://nickmccarty.me/blog/fine-tune-view.html)
- [The OSINT Skill: 11-Layer Target Enrichment for Research Tasks](https://nickmccarty.me/blog/osint-skill.html)
- [The Submit View: Queuing Tasks and Watching the Pipeline Execute in Real Time](https://nickmccarty.me/blog/submit-view.html)
- [The System View: Governance Docs, Active Configuration, and 38 Skills](https://nickmccarty.me/blog/system-view.html)
- [The Analytics View: Score Trends, Token Spend, and Run Distribution](https://nickmccarty.me/blog/analytics-view.html)
- [The Autoresearch View: 40 Experiments, One Keep, and a 3% Signal Rate](https://nickmccarty.me/blog/autoresearch-view.html)
- [The Runs View: Pipeline Stage Visualization and Live Run Monitoring](https://nickmccarty.me/blog/runs-view.html)
- [The GitHub View: Repo Health at a Glance](https://nickmccarty.me/blog/github-view.html)
- [The Memory View: 2,173 Observations, Quality Signals, and an Ontology Graph](https://nickmccarty.me/blog/memory-view.html)
- [Harness vs. Perplexity: Four Iterations to a Tie](https://nickmccarty.me/blog/perplexity-vs-harness.html)
- [The Alt-Data Pipeline: From Beige Book to Paper Trading Thesis](https://nickmccarty.me/blog/alt-data-pipeline.html)
- [Small Language Models and the Efficiency-Accuracy Frontier](https://nickmccarty.me/blog/slm-efficiency.html)
- [Fine-Tuning and Alignment Deep Cuts: Synthetic Data, Poisoning, and Safety Recovery](https://nickmccarty.me/blog/alignment-fine-tuning.html)
- [Knowledge Graphs and Structured Extraction: Beyond the Vector Store](https://nickmccarty.me/blog/knowledge-graphs-extraction.html)
- [Prompt Injection and Agentic Security: What the Attack Literature Says](https://nickmccarty.me/blog/prompt-injection-security.html)
- [Agentic Tool Use and Planning: What the Literature Says](https://nickmccarty.me/blog/tool-use-planning.html)
- [Automated Evaluation Robustness: Metamorphic Testing, Scoring Bias, and Prompt Sensitivity](https://nickmccarty.me/blog/evaluation-robustness.html)
- [Evaluation Uncertainty, Calibration, and Harness Reliability](https://nickmccarty.me/blog/evaluation-uncertainty.html)
- [Judge Benchmarks and Test-Time Scaling: Where LLM Judges Succeed and Where They Don't](https://nickmccarty.me/blog/judge-benchmarks-scaling.html)
- [Judging the Judges: Benchmark Contamination and Evaluation Reliability](https://nickmccarty.me/blog/judge-reliability.html)
- [Structured Knowledge Queries: Ontology, SPARQL, and Grounded Verification](https://nickmccarty.me/blog/structured-knowledge-queries.html)

### Analysis

- [Leverage: What the Metric Measures, What It Doesn't, and Why the Replacement Framing Gets the Math Wrong](https://nickmccarty.me/blog/leverage.html)

### Computer Vision

- [Layer 4 Circuit Diagram](https://nickmccarty.me/blog/layer4-circuit-diagram.html)
- [Circuit Extraction Notebook](https://nickmccarty.me/blog/circuit-extraction-notebook.html)
- [Circuit Extraction for Pot Detection](https://nickmccarty.me/blog/circuit-extraction-pot-detection.html)
- [Sparse Linear Probing for Efficient Detection](https://nickmccarty.me/blog/sparse-probing-pot-detection.html)
- [Extracting Features from Vision Model Backbones](https://nickmccarty.me/blog/extracting-vision-features.html)
- [Mechanistic Interpretability for Agricultural AI](https://nickmccarty.me/blog/mechanistic-interpretability.html)
- [SAM vs Faster R-CNN: A Practical Comparison](https://nickmccarty.me/blog/sam-vs-fasterrcnn-comparison.html)
- [Fine-Tuning Vision Foundation Models](https://nickmccarty.me/blog/fine-tuning-vision-models.html)

### Geospatial AI

- [Object Detection on Drone Orthomosaics with SAM](https://nickmccarty.me/blog/sam-object-detection.html)
- [Building a GeoTIFF Object Detection Web App](https://nickmccarty.me/blog/fastapi-web-app.html)
- [Training Faster R-CNN for Geospatial Object Detection](https://nickmccarty.me/blog/training-faster-rcnn-geospatial.html)

### Other

- [harness-data-model](https://nickmccarty.me/blog/harness-data-model.html)
- [Seven Principles and a Moving Frontier: The Harness Roadmap](https://nickmccarty.me/blog/harness-roadmap.html)

## Site notes

- All inference runs locally (no cloud API for the core harness pipeline)
- Posts are generated by the harness itself via /lit-review and research skills,
  then edited and published manually
- Search index at /search-index.json (100 posts, ~740 KB)
- Sitemap at /sitemap.xml