The September 2026 Cache Price War
Five vendors repriced prompt caching in three weeks. Verified cache-read rates and a worked 200-step agent bill that lands 1.8x apart on two identical price cards.
Read the analysis →Reviewed AI Agent Guides
Thirty-five practical guides reviewed for unsupported claims, dated evidence, and reader value. Fast-changing facts link to primary sources; system-design advice states assumptions and limits.
These guides are independent educational material, not vendor endorsements. Each article identifies the responsible editorial desk, its review date, and the review method. You can also report a correction.
What qualifies for this library
Original utility
A guide must add a decision framework, worked calculation, contract, threat model, or reproducible test plan—not merely restate vendor documentation.
Traceable evidence
Changing technical claims link to primary sources. Illustrative numbers and proposed experiments are labeled so readers do not mistake them for production measurements.
Maintained boundaries
Indexed guides disclose review dates, limitations, and correction paths. Draft or unsupported pages stay outside the sitemap until they meet the same standard.
Choose an AI Model & Agent Stack
Turn task requirements, safety gates, evaluations, latency, and total cost into a defensible selection.
Read guide →AI Agent Systems Explained
Build an inspectable agent loop with explicit state, budgets, traces, permissions, and production-shaped evaluations.
Read guide →AI Coding Agent Security
Threat-model secrets, prompt injection, tool permissions, dependencies, approvals, and local deployment.
Read guide →Fine-Tuning vs RAG
Compare retrieval, training, and hybrid systems with a worked decision case, fair test set, cost model, and rollout gates.
Read guide →MCP Architecture & Security
Review the MCP lifecycle, permission manifests, transport boundaries, OAuth risks, adversarial tests, and revocation.
Read guide →Reliable Structured Outputs
Test schemas with a worked contract, typed failures, semantic validation, measurable quality gates, and safe migrations.
Read guide →LLM API Cost Planning
Work through an auditable monthly estimate, sensitivity analysis, trace schema, and forecast validation process.
Read guide →Prompt Caching
Calculate cache break-even, fingerprint reusable prefixes, and run a reproducible cold-versus-warm experiment.
Read guide →Local LLM Deployment
Run a reproducible Ollama, llama.cpp, and vLLM comparison with explicit artifacts, metrics, and failure tests.
Read guide →Architecture and orchestration
How to Build an AI Agent
Define a task contract, state, tools, budgets, approvals, and an evaluation harness before choosing a framework.
Read guide →AI Agent Memory
Separate working, episodic, semantic, and procedural memory with explicit write, retrieval, and deletion rules.
Read guide →Agentic RAG
Turn retrieval into an observable state machine with bounded retries, evidence checks, and abstention.
Read guide →LangGraph vs CrewAI
Compare orchestration approaches using the same workflow, failure injections, and operational criteria.
Read guide →DSPy Optimization
Build a reproducible compile-and-evaluate loop with representative examples, metrics, and a held-out set.
Read guide →Multi-Model Workflows
Test routers against a single-model baseline while tracking quality, latency, cost, and fallback behavior.
Read guide →Mixture of Experts Explained
Audit MoE claims by separating total parameters, active parameters, routing behavior, and measured serving results.
Read guide →Evaluation, reliability, and security
When Agents Attack
The PaperCut agent swarm and Google ADK's CVSS 10.0 — two verified September 2026 incidents, and what they change about agent defenses.
Read the incident record →How to Evaluate AI Agents
Score outcomes, trajectories, policy compliance, cost, and reproducibility instead of relying on one pass rate.
Read guide →Agentic Testing
Combine deterministic checks, rubric evaluation, adversarial cases, replay, and production monitoring.
Read guide →Reading SWE-bench Results
Interpret repository task benchmarks with attention to setup, contamination, patch validity, and operational cost.
Read guide →Multimodal Evaluation
Design slice-based image and document evaluations with grounded evidence, repair loops, and accessibility checks.
Read guide →Secure Function Calling
Enforce authorization, validation, least privilege, idempotency, logging, and approval outside the model.
Read guide →RAG Chunking Strategies
Choose chunking with a retrieval experiment that measures answer support, boundary failures, latency, and cost.
Read guide →Prompt Compression
Reduce tokens only after checking answer quality, citation retention, instruction survival, latency, and spend.
Read guide →Developer workflows and deployment
Claude Code Workflow
Use repository instructions, scoped permissions, small changes, tests, and diff review as safety boundaries.
Read guide →Cursor vs Windsurf vs Copilot
Run a controlled IDE-agent trial on your own repository rather than trusting a feature checklist.
Read guide →Claude vs GPT vs DeepSeek
Compare model families with a private task set, documented settings, safety checks, latency, and total cost.
Read guide →AI Agents in GitHub Actions
Design a constrained CI trust boundary for untrusted pull requests, secrets, permissions, artifacts, and approvals.
Read guide →Run DeepSeek-R1 with Ollama
Record the exact artifact, resource limits, prompts, measurements, and security boundary for local experiments.
Read guide →Self-Host DeepSeek-R1
Size hardware from the selected checkpoint and quantization, then validate memory, throughput, quality, and exposure.
Read guide →vLLM OpenAI-Compatible Serving
Treat API compatibility as a tested contract covering schemas, streaming, errors, cancellation, and overload.
Read guide →vLLM vs Ollama
Choose a runtime with a migration pilot measuring deployment effort, quality, latency, throughput, and operations.
Read guide →AI Subscription Limits
Verify current account-specific caps and reset behavior without presenting changing limits as universal facts.
Read guide →How to Verify AI Release Claims
Use a dated evidence ledger to distinguish primary announcements, measured results, marketing claims, and unknowns.
Read guide →AI debate & evidence
Contested AI claims, answered with dated sources instead of opinions. Each review separates what was measured from what was estimated, and states where the evidence conflicts.
Should AI Development Slow Down?
What "pacing the frontier" actually proposes, who supports it, and what the proposed safeguards would cost.
Read review →Is AI Actually Dangerous?
What 1,663 recorded incidents show about misuse, failure, and which harms are actually being reported.
Read review →What a Frontier Model Costs to Train
Training cost by generation, and why the companies selling compute oppose a slowdown.
Read review →Is AI Taking Entry-Level Jobs?
Payroll data and employer surveys tell opposite stories; here is where they agree and why they diverge.
Read review →Interactive tools
Use the five model and agent tools to estimate costs, inspect vendor benchmarks, create a model shortlist, or generate an architecture blueprint.