# Swarm Signal

> Independent AI publication covering agents, multi-agent systems, and the engineering challenges of deploying autonomous AI. 103 articles grounded in arxiv papers, technical reports, and production case studies. Published by Tyler Casey.

This file highlights 50 key articles from 103 total (22 Guides + 81 Signals). For the complete article content, see [llms-full.txt](https://swarmsignal.net/llms-full.txt). For citable statistics and findings, see [llms-ctx.txt](https://swarmsignal.net/llms-ctx.txt).

## Agent Design

Architecture, tool use, orchestration patterns, and failure modes for production AI agents.

- [Computer-Use Agents Can't Stop Breaking Things](https://swarmsignal.net/computer-use-agents-ai-browser-automation-anthropic-computer/): Examines computer-use agent failure modes: unauthorized file deletion, GUI prompt injection, and the trust gap between demos and production.
- [The Control Interface Problem in Physical AI](https://swarmsignal.net/physical-ai-and-embodied-agents-2026-humanoid-robots-vision/): Examines the control interface problem in physical AI. Covers NVIDIA Cosmos, VLA training gaps, spatial reasoning failures, and humanoid robots.
- [Knowledge Graphs Just Made RAG Worth the Complexity](https://swarmsignal.net/graphrag-knowledge-graphs-combined-with-retrieval-augmented/): Evaluates GraphRAG architectures combining knowledge graphs with RAG. Analyzes multi-hop reasoning improvements and implementation cost trade-offs.
- [The Observability Gap in Production AI Agents](https://swarmsignal.net/ai-agent-observability-and-monitoring-in-production-distribu/): Evaluates the observability gap in production AI agents. Covers reasoning vs execution anomalies, distributed tracing, and audit systems.
- [Function Calling Is the Interface AI Research Forgot](https://swarmsignal.net/function-calling-and-tool-use-in-llms-how-ai-agents-interact/): Examines the gap between function calling capability and production tool use reliability. Covers parameter extraction, RC-GRPO, and multilingual failures.
- [AI Agents Are Security's Newest Nightmare](https://swarmsignal.net/prompt-injection-attacks-on-ai-agents-indirect-prompt-inject/): Evaluates prompt injection attack vectors against AI agents. Covers MUZZLE benchmark (92% hijack rate), CausalArmor defenses, and mitigation.
- [When AI Agents Have Tools, They Lie More](https://swarmsignal.net/ai-agent-hallucinations-why-agents-hallucinate-with-tool-acc/): Examines why tool-augmented AI agents hallucinate more than standard chatbots. Covers trajectory drift, fabricated tool outputs, and mitigation.
- [Why Agent Builders Are Betting on 7B Models Over GPT-4](https://swarmsignal.net/small-language-models-slms-vs-llms-for-ai-agents-efficient-o/): Compares small language models vs frontier LLMs for agent tasks. Analyzes cost-performance with Gemma 2 9B, Phi-3-mini, and quantized models.
- [When Your Judge Can't Read the Room](https://swarmsignal.net/llm-as-judge-ai-evaluation-at-scale-pointwise-scoring-pairwi/): Evaluates LLM-as-Judge systems for AI evaluation at scale. Covers pointwise scoring, pairwise comparison, position bias, and accuracy limits.
- [Types of AI Agents: Reactive, Deliberative, Hybrid, and What Comes Next](https://swarmsignal.net/types-of-ai-agents/): Classifies AI agents into four types (reactive, deliberative, hybrid, autonomous) with production tradeoffs and real-world deployment examples.
- [How to Test and Debug AI Agents](https://swarmsignal.net/testing-debugging-ai-agents/): Guide to testing AI agents covering the MAST failure taxonomy (14 modes), CLEAR consistency framework, and a production testing pipeline.
- [From Prompt to Partner: A Practical Guide to Building Your First AI Agent](https://swarmsignal.net/from-prompt-to-partner-a-practical-guide-to-building-your-first-ai-agent/): Practical guide to building production AI agents. Covers the three pillars (model, tools, instructions), ReAct framework, and deployment patterns.
- [Hierarchical Agents Don't Know Who They're Talking To](https://swarmsignal.net/hierarchical-agents-dont-know-who-theyre-talking-to/): Analyzes the personalization gap in hierarchical multi-agent systems. Covers how user preference signals degrade through each layer of agent stacks.
- [When Your Agent Stops Using Tools](https://swarmsignal.net/when-your-agent-stops-using-tools/): Examines "interaction collapse" in RL-trained agents that progressively abandon tool use mid-session. Covers ASTER and CM2 diagnostic frameworks and fixes.
- [Agent System Prompts Are Untested](https://swarmsignal.net/agent-system-prompts-are-untested/): Documents how production agent system prompts ship without systematic testing. Analyzes the gap between prompt engineering and validated prompt assurance.
- [Your Agent's Memory Problem](https://swarmsignal.net/your-agents-memory-problem/): Examines memory fragmentation in production agents. Covers context window limitations, working memory decay, and cross-session persistence failures.
- [Agent Verification Gap](https://swarmsignal.net/agent-verification-gap/): Evaluates the lack of formal verification in deployed AI agents. Covers runtime assertion checking, proof-carrying agent outputs, and verification costs.
- [AI Agent Security Playbook](https://swarmsignal.net/ai-agent-security-playbook/): Comprehensive security framework for AI agents. Covers attack surface mapping, defense-in-depth architectures, and incident response for agent-specific threats.
- [Agentic AI Coding Assistants: Production Reliability](https://swarmsignal.net/agentic-ai-coding-assistants-production-reliability/): Evaluates production reliability of agentic coding assistants. Covers error rates, rollback patterns, and the gap between demo performance and real-world coding tasks.

## Swarm Systems

Multi-agent coordination, swarm intelligence, communication protocols, and collective behavior.

- [Agents That Reshape, Audit, and Trade With Each Other](https://swarmsignal.net/agents-that-reshape-audit-and-trade-with-each-other/): Analyzes emerging multi-agent behaviors: dynamic topology reconfiguration, self-auditing deception detection, and economic negotiation.
- [AI Agent Orchestration Patterns: From Single Agent to Production Swarms](https://swarmsignal.net/ai-agent-orchestration-patterns/): Taxonomy of six production agent orchestration patterns with framework implementations (CrewAI, LangGraph, AutoGen, OpenAI Agents SDK).
- [Swarm Intelligence Explained: From Ant Colonies to AI Agent Fleets](https://swarmsignal.net/swarm-intelligence-explained/): Traces swarm intelligence from ant colonies through classical algorithms (ACO, PSO) to modern LLM agent swarms and production applications.
- [Multi-Agent Systems Explained: How AI Agents Coordinate, Compete, and Fail](https://swarmsignal.net/multi-agent-systems-explained/): Guide to multi-agent systems covering coordination patterns, competition dynamics, failure modes, and production orchestration strategies.
- [AI Agents Built a Social Network](https://swarmsignal.net/ai-agents-built-a-social-network/): Documents autonomous agents forming emergent social structures when deployed in unstructured environments. Analyzes agent-to-agent trust networks and coordination.
- [Multi-Agent Coordination Failures](https://swarmsignal.net/multi-agent-coordination-failures/): Catalogues coordination failure modes in multi-agent deployments. Covers task allocation deadlocks, information bottlenecks, and cascading errors.
- [LLM Swarm 300x Problem](https://swarmsignal.net/llm-swarm-300x-problem/): Quantifies the 300x cost multiplier when naive agent swarms multiply LLM calls. Analyzes token overhead, redundant reasoning, and cost optimization strategies.

## Reasoning & Memory

RAG architectures, agent memory, context engineering, inference-time compute, and reasoning tokens.

- [From Goldfish to Elephant: How Agent Memory Finally Got an Architecture](https://swarmsignal.net/agent-memory-architecture-guide/): Guide to agent memory architectures covering tiered memory (MemGPT), budget-aware allocation, and temporal knowledge graphs.
- [From Answer to Insight: Why Reasoning Tokens Are a Quiet Revolution in AI](https://swarmsignal.net/from-answer-to-insight-why-reasoning-tokens-are-a-quiet-revolution-in-ai/): Explains reasoning tokens (hidden chain-of-thought) comparing OpenAI o1, DeepSeek R1, and Claude extended thinking with cost-performance data.
- [The Goldfish Brain Problem: Why AI Agents Forget and How to Fix It](https://swarmsignal.net/the-goldfish-brain-problem-why-ai-agents-forget-and-how-to-fix-it/): Guide to the AI agent memory problem. Covers Stanford's generative agents, MemGPT's tiered architecture, and three policy approaches to persistence.
- [Multi-Agent Reasoning's Memory Problem](https://swarmsignal.net/multi-agent-reasonings-memory-problem/): Examines how reasoning models fail to surface their own parametric knowledge. Stanford study shows 40% knowledge-access failure rate in RL-trained models.
- [Your Model Already Knows the Answer](https://swarmsignal.net/your-model-already-knows-the-answer/): Explores latent knowledge access in LLMs. Analyzes the gap between what models encode and what they surface during reasoning.

## Safety & Governance

AI safety reports, alignment, bias, regulation (EU AI Act), benchmarks, and red teaming.

- [AI Guardrails for Agents: How to Build Safe, Validated LLM Systems](https://swarmsignal.net/ai-guardrails-agents/): Compares four guardrail systems (NeMo, Guardrails AI, Bedrock, Llama Guard) and proposes a five-layer production architecture for agent safety.
- [Nobody Knows If Deployed AI Agents Are Safe](https://swarmsignal.net/nobody-knows-if-deployed-ai-agents-are-safe/): Evaluates the absence of safety standards for production AI agents. Covers missing benchmarks, untested failure modes, and the gap between lab safety and deployment reality.
- [One Fake Source Broke Every Agent](https://swarmsignal.net/one-fake-source-broke-every-agent/): Documents how a single poisoned information source cascades through multi-agent systems. Analyzes trust propagation failures and content verification gaps.
- [Alignment Backfires in Japanese](https://swarmsignal.net/alignment-backfires-in-japanese/): Examines how AI alignment techniques produce unintended behavior in non-English contexts. Covers cultural bias in RLHF and cross-lingual safety failures.
- [Weaponized Swarms and Democracy](https://swarmsignal.net/weaponized-swarms-democracy/): Analyzes the threat of coordinated AI agent swarms to democratic processes. Covers information operations, synthetic media distribution, and countermeasures.

## Models & Frontiers

Frontier model comparisons, MoE architectures, training data, deployment gaps, and open-weight models.

- [Inference-Time Compute Is Escaping the LLM Bubble](https://swarmsignal.net/inference-time-compute-scaling/): Reviews inference-time compute scaling expanding beyond LLMs into diffusion models and code generation. Covers UnMaskFork and reward model routing.
- [Synthetic Data Won't Save You From Model Collapse](https://swarmsignal.net/synthetic-data-generation-for-ai-training-model-collapse-whe/): Reviews synthetic data risks including model collapse, domain-specific failures in medical imaging, and mitigation strategies for training pipelines.
- [MoE Models Run 405B Parameters at 13B Cost](https://swarmsignal.net/mixture-of-experts-architecture-sparse-moe-expert-routing-in/): Reviews sparse MoE architecture from routing mechanisms to load balancing failures. Analyzes DeepSeek-V3, Qwen-2.5-MoE, and deployment costs.
- [Mixture of Experts Explained: The Architecture Behind Every Frontier Model](https://swarmsignal.net/mixture-of-experts-explained/): Explains MoE architecture from 1991 origins through modern implementations (DeepSeek-V3, Mixtral, Switch Transformer). Covers routing and inference.
- [FlashAttention 4 and Asymmetric Scaling](https://swarmsignal.net/flashattention-4-asymmetric-scaling/): Evaluates FlashAttention 4's asymmetric scaling properties. Covers hardware-aware optimization, memory bandwidth constraints, and implications for long-context inference.
- [MoE Training Just Got 4x Faster](https://swarmsignal.net/moe-training-just-got-4x-faster/): Reports on MoE training speedups through improved expert parallelism. Covers load balancing innovations, communication overhead reduction, and scaling benchmarks.

## Real-World AI

Enterprise deployment, national AI strategies, drug discovery, coding productivity, and workforce impact.

- [AI Coding Assistants: The Productivity Paradox](https://swarmsignal.net/ai-coding-productivity-paradox/): Documents the disconnect between individual AI coding productivity gains and organizational delivery outcomes using Faros AI data.
- [AI in Drug Discovery: From Hype to Clinical Proof](https://swarmsignal.net/ai-drug-discovery/): Tracks AI in pharmaceutical development from experimental tool to clinical-stage reality. Covers AlphaFold, Insilico Medicine, and Recursion.
- [Obsidian's CLI Turns Your Second Brain Into an API](https://swarmsignal.net/obsidian-cli-guide/): Hands-on guide to Obsidian 1.12's CLI with 100+ commands. Covers setup, AI agent integration, and the architecture enabling personal knowledge APIs.
- [AI in Customer Service: Enterprise Autonomous Support](https://swarmsignal.net/ai-agents-in-customer-service-and-enterprise-autonomous-supp/): Evaluates AI agent deployment in customer service. Covers resolution rates, escalation patterns, customer satisfaction, and cost reduction metrics.
- [The Trillion Dollar Agent Panic](https://swarmsignal.net/trillion-dollar-agent-panic/): Documents enterprise panic around AI agent adoption. Covers competitive pressure, rushed deployments, and the gap between strategic intent and operational readiness.
- [LLM Agents Can't Handle Markets](https://swarmsignal.net/llm-agents-cant-handle-markets/): Examines AI agent failures in market-based environments. Covers price discovery, strategic deception, and the difficulty of agent participation in competitive markets.
- [AI Agents in Japan](https://swarmsignal.net/ai-japan/): Surveys Japan's AI agent landscape. Covers robotics integration, cultural factors in agent adoption, and regulatory frameworks specific to autonomous systems.
- [AI Agents in China](https://swarmsignal.net/ai-china/): Analyzes China's AI agent ecosystem. Covers open-source model adoption, multi-agent platform competition, and policy-driven deployment strategies.

## Optional

- [About Swarm Signal](https://swarmsignal.net/about/): Publication background, editorial approach, and author credentials
- [Full Article Content](https://swarmsignal.net/llms-full.txt): Complete text of all 103 articles for direct ingestion
- [Citation Index](https://swarmsignal.net/llms-ctx.txt): All citable statistics, findings, and question-to-article lookup
- [RSS Feed](https://swarmsignal.net/rss/): Full-text RSS feed for all articles
- [Sitemap](https://swarmsignal.net/sitemap.xml): XML sitemap for all published content

> Last updated: 2026-03-13