Insights
How to build, run and secure AI agents in production.
Practical guides for engineers and technology leaders taking AI agents from pilot to production: context and loops, cost, evaluation and operations, identity and security. Each guide is checked against primary sources on the day it is written and says where its advice stops. Where a guide compares platforms, it shows Anthropic, OpenAI, Google Cloud, AWS and Microsoft side by side.
Evaluation · · 13 min read
AI Agent Benchmark Scores: What They Actually Prove
An agent benchmark score can be a broken test. Six questions to ask before you trust one, the metrics to ask for beside it, and where public scores stop.
Context and memory · · 19 min read
Context Engineering for AI Agents: What the Model Reads
Context engineering is deciding what an AI agent reads on every model call: instructions, tools, retrieval, memory, tool results and history, and in what order.
Agent engineering · · 19 min read
Harness Engineering: Why the Same Model Scores Differently
Harness engineering is the design of everything around the model: tools, sandbox, permissions, state and hooks. What it is, and where each control stops.
Cost and FinOps · · 23 min read
Token Optimization for AI Agents: What to Fix First
An agent's bill is mostly re-reading. Cache the stable prefix, send less, set reasoning per task, route whole tasks, batch what can wait. A worked cost model.
Cost and FinOps · · 20 min read
FinOps for AI Agents: Cost per Task, Not per Token
FinOps for AI agents means pricing each completed task, not each token: a meter you build, reconciled daily with the bill, plus budgets, caps and showback.
AgentOps · · 18 min read
AgentOps: How to Operate AI Agents in Production
AgentOps is operating AI agents against quality, not only uptime: version the whole agent, gate releases on evals, canary, trace with OpenTelemetry, roll back.
Agent engineering · · 16 min read
Graph Engineering for AI Agents: The Step That Runs Twice
Graph engineering builds an AI agent as explicit nodes, edges and state, with checkpoints and interrupts in ADK and LangGraph, and every step safe to run twice.
Agent engineering · · 19 min read
Loop Engineering: How to Design Agent Loops That Stop
Loop engineering decides, in code, what starts an AI agent, what proves it is done and what stops it: turn, token and time limits in ADK, Claude and OpenAI.
Agent security · · 17 min read
AI Agent Identity: Who Acts, and on Whose Behalf
Give each AI agent its own identity, use delegation not impersonation, and keep tokens narrow. Patterns for Google Cloud, AWS, Microsoft, Anthropic and OpenAI.
Production readiness · · 14 min read
AI Agent Registry: Count Your Agents Before You Govern Them
An AI agent registry lists your agents, MCP servers and tools with an owner, an identity and allowed calls. A gateway enforces it and refuses by default.
- Research
Evaluation · · 7 min read
Checking AI Docs for Hallucinations Without an LLM Judge
Check AI-written docs for hallucinations in code: resolve every file path, name and citation against the repository, list what fails, and test the checker.
Agent security · · 9 min read
Human in the Loop for AI Agents: Four Autonomy Tiers
Human in the loop for AI agents: a person approves each consequential action until recorded evidence earns more autonomy. Four tiers, from assist to high.
Agent engineering · · 5 min read
MCP vs A2A: What Each Protocol Does for AI Agents
MCP connects an agent to tools and data; A2A connects agents to other agents. What each protocol defines, where they meet, and when a system needs both.
New guides by email
Get an email when we publish a new guide on building, running and securing AI agents. You can unsubscribe from any of our emails.
Signing up needs JavaScript, and with scripts switched off this form cannot send. You can follow new guides through the feed instead. Insights feed (RSS)
Building agents like these?
Aelsha Innovations designs, builds, secures and observes AI agents in production. A readiness assessment is the place to start.