Less hype. More working systems.
AgentEngineering.org covers the design, tooling, evals, failure modes, and operating practice behind AI agents that need to survive real work.
Start With These Foundations
Start Here
Frame the discipline before the reader enters the foundations sequence.
Start here What Is Agent Engineering?Phase 1: Foundations of Agency
Define what an agent is, what changes from plain LLM use, and how autonomy should be understood.
Start here What Is an AI Agent?Phase 2: Core Mechanics
Show how agent systems decompose work, take action, remember, and reason through multi-step runs.
Start here Planning and Task DecompositionStart Here First
See the full archiveIf you landed here from search, start with these cornerstone guides before dropping into the full archive. They define the site’s main vocabulary, design choices, and production operating model.
What Is Agent Engineering?
The discipline of designing, evaluating, and operating goal-directed AI systems that use tools and act under explicit control.
What Is an AI Agent?
An AI agent is a goal-directed system that can observe state, decide what to do next, use tools, and act across multiple steps.
LLMs, Workflows, and Agents: What Actually Changes?
The real shift from LLM to workflow to agent is not a buzzword change. It is a change in who owns the task, the execution path, and the next-step decisions.
When to Use a Workflow Instead of an Agent
Use a workflow when the valid path can be defined in advance, predictability matters more than flexibility, and the task does not need runtime path-finding.
Tool Use: How Agents Take Action
Tool use is how an agent leaves pure text generation and interacts with external systems. Reliable tool use depends on more than choosing a function name. It depends on arguments, execution control, permissions, and verification.
Structured Outputs, Guardrails, and Execution Boundaries
Structured outputs constrain shape, guardrails constrain policy, and execution boundaries constrain power. Safe agent systems need all three.
Tracing and Observability for Agent Systems
An uninstrumented agent run looks like an HTTP 200. OpenTelemetry GenAI conventions (still Development), span types, a working example, and the tool landscape.
AgentOps: Running Agents in Production
AgentOps is the operating discipline for live agent systems. It turns traces, evaluations, guardrails, and human controls into an ongoing practice for running autonomous systems safely and reliably.
AI Agent Frameworks
Nine agent frameworks: versions from each project's docs, a two-axis map, and the same task in LangGraph, OpenAI Agents SDK, Claude Agent SDK, and Google ADK.
Tool Integration Patterns for Real Agent Systems
Tool integration is a durable agent design problem about boundaries, trust, and execution control. MCP matters, but it is one interface pattern inside a much larger tool story.
Section Hubs
See the full learning pathOpinions
Founder-led arguments, market reads, and point-of-view essays about where agent engineering is actually heading.
Tools
Coverage focused on the surfaces agents touch directly: tool calling, context pipelines, memory, schemas, guardrails, and execution boundaries.
Platforms
Agent SDKs, runtimes, orchestration products, and build surfaces you can use to create custom agents.
Featured Articles
Error Analysis: The First Eval You Should Run
Read traces and count what breaks before writing an eval. Open coding, axial coding, and saturation decide which code assertions and binary judges to build.
Agent Security: Prompt Injection and the OWASP Agentic Top 10
Prompt injection has no patch, so the job is containment. This guide covers why the data-instruction boundary does not exist, the lethal trifecta, how the 2026 OWASP LLM and Agentic lists divide the ground, the ten agentic risks grouped for builders, defenses ranked by how much they actually guarantee, and runnable code for taint tracking, output sanitising, and tool-description pinning.
MCP in Practice (2026-07-28 Spec)
The Model Context Protocol went stateless in the 2026-07-28 revision. This guide covers the architecture, the three server primitives, both transports, authorization, a small working server, the security model, production operations, and when not to use MCP at all.
All Articles
Error Analysis: The First Eval You Should Run
Read traces and count what breaks before writing an eval. Open coding, axial coding, and saturation decide which code assertions and binary judges to build.
Agent Security: Prompt Injection and the OWASP Agentic Top 10
Prompt injection has no patch, so the job is containment. This guide covers why the data-instruction boundary does not exist, the lethal trifecta, how the 2026 OWASP LLM and Agentic lists divide the ground, the ten agentic risks grouped for builders, defenses ranked by how much they actually guarantee, and runnable code for taint tracking, output sanitising, and tool-description pinning.
MCP in Practice (2026-07-28 Spec)
The Model Context Protocol went stateless in the 2026-07-28 revision. This guide covers the architecture, the three server primitives, both transports, authorization, a small working server, the security model, production operations, and when not to use MCP at all.
Designing Tools for Agents: The Agent-Computer Interface
Tool definitions are the interface an agent actually programs against. This guide covers which tools to build, how to name and describe them, parameter design that prevents bad calls, response shaping, actionable errors, safe repeats, and how to evaluate a tool surface.
How to Review an AI Agent Demo Without Getting Fooled
A 30-minute AI agent demo can prove or disprove production readiness if you know what to test live, what to ask the builder, and what to refuse to accept as proof. The D.E.M.O. lens gives you four tells.
Tool Integration Patterns for Real Agent Systems
Tool integration is a durable agent design problem about boundaries, trust, and execution control. MCP matters, but it is one interface pattern inside a much larger tool story.
AI Agent Frameworks
Nine agent frameworks: versions from each project's docs, a two-axis map, and the same task in LangGraph, OpenAI Agents SDK, Claude Agent SDK, and Google ADK.
The Most Common Ways Agents Fail Silently
The most dangerous agent failures are often not dramatic incidents. They are quieter losses of trust: acceptable-looking outputs hiding weaker trajectories, more rescue, noisier grounding, and rising pressure on the system's real operating limits.
Traces as Test Data: Using Production Runs to Improve Agent Quality
Production traces are not just for debugging. The best ones become future quality protection: regression fixtures, scenario cases, and stronger offline evals. The trick is knowing which traces deserve promotion.
Online Evals vs Offline Evals
Offline evals decide whether a change deserves release. Online evals judge how the live system is actually behaving under real traffic. Production agent teams need both, and they need them for different reasons.
What Is Agent Engineering?
The discipline of designing, evaluating, and operating goal-directed AI systems that use tools and act under explicit control.
How Good Agent Memory Actually Works in Production
Good agent memory is not one vector store plus chat history. It is a governed system for deciding what gets scoped, promoted, compressed, pinned, and retrieved.
Agent Memory Is Growing Up - Why Agents Are Starting to Remember How, Not Just What
Agent memory is changing fast. The next wave of agents will not just remember facts. They will remember workflows, compress experience, and get better at solving the next problem.
Reliability Reviews for Agents
Regression tests protect the next release. Reliability reviews ask a broader question: is this live agent system still trustworthy enough to keep operating as designed?
Regression Testing for Agents
Regression testing is the release-gate discipline that checks whether an agent got worse after a change. For agent systems, that means testing not only outputs, but also trajectories, side effects, and operating envelopes.
AgentOps: Running Agents in Production
AgentOps is the operating discipline for live agent systems. It turns traces, evaluations, guardrails, and human controls into an ongoing practice for running autonomous systems safely and reliably.
Tracing and Observability for Agent Systems
An uninstrumented agent run looks like an HTTP 200. OpenTelemetry GenAI conventions (still Development), span types, a working example, and the tool landscape.
OpenAI Codex as a Coding-Agent Platform
Codex is one open-source harness behind many clients, reachable over a JSON-RPC protocol. What that buys you, and where it ties you in.
Evaluating Agent Trajectories, Not Just Outputs
A correct final answer does not prove that an agent behaved well. Agent evaluation has to judge the run itself: the sequence, tool use, recovery behavior, and policy fit that produced the answer.
Human-in-the-Loop Control Design
Human-in-the-loop design is not about adding vague oversight. It is about deciding where human judgment should sit in an agent system and what type of checkpoint belongs there.
Supervisor, Router, and Planner-Executor Patterns
Routers dispatch, planners break work into a roadmap, and supervisors retain control across the run. The right orchestration pattern depends on where authority should live.
Structured Outputs, Guardrails, and Execution Boundaries
Structured outputs constrain shape, guardrails constrain policy, and execution boundaries constrain power. Safe agent systems need all three.
When to Use a Workflow Instead of an Agent
Use a workflow when the valid path can be defined in advance, predictability matters more than flexibility, and the task does not need runtime path-finding.
ReAct and the Basic Reasoning Loop
ReAct interleaved reasoning with actions and became the default agent loop. The idea survived; the parsed Thought/Action/Observation format did not.
Goals, Constraints, and Success Conditions
Goals tell an agent what outcome to pursue. Constraints define the boundaries on how it may pursue that outcome. Success conditions define what evidence lets the run stop. Real agents need all three.
The Autonomy Spectrum: From Stateless Calls to Goal-Directed Systems
Autonomy is not a binary property that suddenly appears when a system uses tools or takes multiple steps. It is a spectrum shaped by who chooses goals, path, actions, and recovery behavior at runtime.
Context Engineering: The New Core Skill
Context engineering is not a replacement for prompt engineering. It is a specialization inside prompt engineering focused on constructing the dynamic, system-heavy parts of the final prompt payload.
Short-Term Context, Retrieval, and Long-Term Memory
Agents do not just need more context. They need clean separation between what the model sees now, what the system can fetch now, and what the system should still know later.
What Stripe's Minions Reveal About Production Coding Agents
Stripe's Minions matter because they show what coding agents look like when they are treated as delegated workers inside a real engineering system. This case study extracts the reusable architecture patterns and compares Stripe's model with Devin and Claude Code.
Tool Use: How Agents Take Action
Tool use is how an agent leaves pure text generation and interacts with external systems. Reliable tool use depends on more than choosing a function name. It depends on arguments, execution control, permissions, and verification.
Planning and Task Decomposition
Planning chooses the path toward a goal. Task decomposition turns that path into executable, verifiable subtasks. In agent systems, the quality of that breakdown often determines whether the run succeeds.
The Sense-Think-Act Loop
The sense-think-act loop is the runtime pattern that makes an AI agent agentic. It turns goals and changing state into repeated bounded actions instead of one-shot responses.
LLMs, Workflows, and Agents: What Actually Changes?
The real shift from LLM to workflow to agent is not a buzzword change. It is a change in who owns the task, the execution path, and the next-step decisions.
What Is an AI Agent?
An AI agent is a goal-directed system that can observe state, decide what to do next, use tools, and act across multiple steps.
Get the biweekly issue
What changed in agent engineering, what it means for systems in production, and one thing to try. Every other Thursday.