← Back to home

Blog — all articles

Every opsagent article on agent security, MCP, governance, observability, evals, orchestration, and oversight — newest first.

Orchestration

What is an AI agent? A definition without the marketing

An AI agent is software that plans and executes multi-step tasks on its own: model, tools, memory, and an action loop. How it differs from a chatbot and from automation, what it's good for, and which risks to control from day one.

Read →
Governance

Policy Orchestration for AI Agents: One Decision Layer, Not Rules Scattered Across Prompts

Rules spread across prompts drift and leave no audit trail. How a policy decision point and enforcement points at the tool boundary give AI agents one rulebook.

Read →
Orchestration

Agentic engineering: building software with coding agents you can actually ship

What agentic engineering means in 2026 — developing software with coding agents like Claude Code, Codex, Cursor and Aider — where the line to vibe coding runs, what AGENTS.md and CLAUDE.md are for, what the METR study measured, and what has to be true after the merge for the result to survive production.

Read →
Security

Indirect prompt injection: detecting and defending against it in your agent pipeline

How to detect and contain indirect prompt injection in a production agent pipeline — content isolation, provenance, output filtering and the layered defenses that shrink the blast radius.

Read →
Security

Red teaming AI agents: building a red-team program in your organization

How to stand up an agent red-team program as a repeatable process — scope, cadence, roles, and how organizational red teaming of autonomous agents differs from a one-off pentest.

Read →
Security

Memory poisoning: defending an agent's memory in production

How to defend a production agent's long-term memory and retrieval store against poisoning — validation, provenance, segmentation and expiry that stop false data biasing future decisions.

Read →
Security

The OWASP Top 10 for Agentic Applications: mitigating each risk in production

A production playbook for the OWASP Top 10 for Agentic Applications: how to meet and mitigate each of the ten risks with concrete controls, guardrails and governance.

Read →
Security

Tool poisoning: when an agent's tools turn against it

How tampered tool descriptions and crafted tool outputs steer autonomous agents into harmful actions — and the validation, isolation and provenance controls that stop it.

Read →
Governance

The EU AI Act timeline for autonomous agents

When each EU AI Act obligation actually bites — the Article 50 transparency duties have applied since 2 August 2026, and the next hard deadline for agent teams is the high-risk requirements on 2 December 2027.

Read →
Security

Excessive agency: the most damaging agent risk

Why over-broad permissions turn a small agent mistake into a major incident — and how least privilege, scoped tools and approval gates contain the blast radius.

Read →
Governance

EU AI Act vs GDPR: two regimes your agents must satisfy

Why GDPR compliance does not cover the EU AI Act, where the two overlap, and what an autonomous agent has to satisfy under each.

Read →
Governance

Writing a company AI policy that actually governs agents

A practical template for an internal AI policy — the sections that matter, how to cover autonomous agents and Shadow AI, and how to keep it from becoming shelfware.

Read →
Governance

ISO 42001 for agents: an AI management system, step by step

What ISO/IEC 42001 requires, how it complements the EU AI Act, and a practical path to an AI management system that actually governs autonomous agents.

Read →
MCP & Identity

MCP security: secure configuration, identity and gateway authorization

How to run the Model Context Protocol safely in production: an MCP gateway, authorization and scoping, and non-human identity controls that keep agent tooling under control.

Read →
Observability

Agent observability: the five signals and the alert rules that watch them

What to measure on a production AI agent — outcome rate, cost per run, loop depth, tool errors and liveness — with copyable alert rules, the OpenTelemetry GenAI conventions that standardise the data, and two failures from operating a real portfolio.

Read →
MCP & Identity

Non-human identity: governing the agents that act for you

Agents authenticate as machine identities, not people. Here is why non-human identity is now a core control — and how to inventory, scope and rotate it.

Read →
MCP & Identity

The OWASP MCP Top 10: securing the protocol agents run on

A practical walkthrough of the OWASP MCP Top 10 — the security risks specific to Model Context Protocol servers — and the controls that address each.

Read →
MCP & Identity

Zero-trust for AI agents: implementing never-trust, always-verify in production

How to operationalize zero-trust for autonomous agents and the tools and agents they call — identity, policy enforcement and segmentation that make continuous verification real in production.

Read →
Observability

Agent cost control: FinOps for autonomous AI

Why cost is a first-class signal for agents, what actually drives the bill, and how token tracking, loop detection and budgets keep an autonomous system affordable.

Read →
Evals

Evaluating AI agents: from evals to reliable production

How to evaluate autonomous agents you can trust — LLM-as-a-judge done right, the metrics that matter, and the CI/CD quality gates that keep agents from regressing.

Read →
Observability

The production metrics that tell you an agent is working

Six metrics — completion, accuracy, hallucination, latency, cost and satisfaction — that turn an agent from a black box into a system you can actually run.

Read →
Observability

Tracing AI agents with OpenTelemetry: spans, attributes and working code

How to instrument an agent with the OpenTelemetry GenAI conventions — invoke_agent and execute_tool spans, gen_ai.* attributes, a Python example you can run — and why a trace with a correlation id once saved this portfolio from rotating a healthy credential.

Read →
Evals

Agent benchmarks: what SWE-bench, τ-bench and GAIA actually measure

The agent benchmarks worth knowing — SWE-bench Verified, τ-bench with its pass^k reliability metric, GAIA's human-vs-model gap — where leaderboard numbers mislead, and a reading checklist for the next score a vendor shows you.

Read →
Orchestration

Agent memory architecture: short-term, long-term, episodic

How autonomous agents remember — the three kinds of memory, when to use each, and why governing the memory layer is both a capability and a security decision.

Read →
Orchestration

Agent orchestration: LangGraph vs CrewAI vs AutoGen and beyond

A vendor-neutral comparison of the main agent frameworks — LangGraph, CrewAI, AutoGen, OpenAI Agents SDK and Google ADK — and how to choose between single-agent and multi-agent designs.

Read →
Evals

Regression testing for AI agents: stop fixing one thing and breaking three

Why a golden dataset and CI gates are the only reliable defense against silent agent regressions when models, prompts and tools change underneath you.

Read →
Evals

LLM-as-a-judge: prompts, code and calibration that make it trustworthy

A working LLM-as-a-judge setup: a judge prompt template you can copy, a DeepEval G-Eval example, the three biases Zheng et al. measured and how to counter them, calibration against human labels, and the cases where a deterministic check beats a judge.

Read →
Human Oversight

Agent guardrails: the automated limits around autonomy

Input, output and behavioral guardrails form the safety layer that lets an agent act fast without acting dangerously. Here is what each does and how they fit together.

Read →
Human Oversight

The agent kill switch: stop conditions, safe state and a drill that proves it

How to design the stop layer for an AI agent — kill switch vs circuit breaker, trip conditions you can copy, what 'halt in a safe state' means under Article 14 of the EU AI Act, and what happens when an automation runs for weeks with no breaker at all.

Read →
Orchestration

Context engineering: giving an agent what it needs, not everything

Why what you put in the context window decides an agent's quality, cost and safety — and the discipline of assembling the right context at each step.

Read →
Human Oversight

Graduated deployment: raising agent autonomy without incidents

A five-stage path — dry-run, read-only, simulation, staging, production — for giving an AI agent more autonomy only after evidence says it is safe.

Read →
Human Oversight

Human oversight and agent autonomy: HITL, HOTL, HOOTL

The spectrum of human oversight for AI agents — from human-in-the-loop to human-out-of-the-loop — plus guardrails, kill switches and a graduated path to raising autonomy safely.

Read →
Human Oversight

Shadow AI: the agents and tools running without your oversight

Why unsanctioned AI use is the most common governance gap — and how to bring Shadow AI into the light without driving it further underground.

Read →
Orchestration

Single-agent or multi-agent? A decision guide

When to split a task across multiple AI agents and when one is enough — the real costs of multi-agent systems and a simple test for whether you actually need one.

Read →