Ideas for building reliable digital systems.

Harness engineering for AI agents

The operational layer that connects goals, tools, memory and controls to make an agent governable.

Loop engineering: reliable cycles for AI agents

An explicit cycle of plan, action and evaluation keeps autonomy from turning into unproductive repetition.

Context and memoryRead article

Context engineering: the right information at the right time

Selecting, structuring and updating relevant context improves agents’ accuracy, cost and traceability.

Tools and controlsRead article

Tool engineering: tools an agent can use safely

Tools are not neutral extensions: they require contracts, validation and interpretable results.

Context and memoryRead article

Memory engineering: useful memory without piling up noise

An agent’s memory needs policies for writing, retrieval, retention and forgetting.

Orchestrating multi-agent systems

Coordination turns specialized agents into a system that can split, verify and integrate complex work.

Tools and controlsRead article

Guardrails and permissions for AI agents

Autonomy is useful when policies, permissions and limits are applied before an action.

Quality and observabilityRead article

Evaluating AI agents: from demos to operational evidence

Evaluating an agent means reviewing result, process, security and cost under representative scenarios.

Human-in-the-loop: oversight that improves decisions

Human review works when it is placed at the right point of risk, with evidence and actionable options.

Quality and observabilityRead article

Observability and tracing: making autonomy visible

Traces, metrics and events connect each result to the execution that produced it.

MCP explained: how it connects AI assistants with tools and data

MCP offers a common way to connect AI applications with servers that expose tools and data. Understanding its components helps separate what the protocol defines from the risks of each implementation.

Where the risk really is: client, server, agent, permissions and integrations

Evaluating MCP is not limited to the protocol. The real surface includes who runs each component, what it can see and which actions it can perform.

Indirect prompt injection: how an external response can manipulate an agent

An agent can interpret as instructions content that arrived as data. The defense requires separating authority, limiting capabilities and controlling actions.

Tool poisoning: when a seemingly trustworthy tool hides instructions

A tool can include descriptions or results that steer the model toward unexpected actions. Review must cover behavior, capabilities and changes.

Rug pull: what happens when a tool changes after being approved

Approving a tool should not mean trusting all of its future versions indefinitely. Change control closes that gap.

Excessive permissions: applying least privilege to MCP servers

Least privilege reduces how much an integration can do if an instruction, configuration or component fails.

Cybersecurity8 min

Tokens, OAuth and the “confused deputy”: preventing unauthorized access

When a server acts on behalf of a person toward another service, it must preserve the identity, consent and scope of each authorization.

Cybersecurity8 min

SSRF and local MCP servers: network risks and defense against DNS rebinding

A local server or a tool that fetches URLs can open unexpected access paths. The defense combines network limits, validation and transport controls.

Command injection and input validation: protecting tools that perform actions

Parameters proposed by a model must be treated as untrusted input. Security depends on validating the type, scope and effect of each operation.

Cybersecurity8 min

Supply chain and unapproved MCP servers: dependencies, installation and shadow IT

Each MCP server adds code, configuration and dependencies to an organization’s supply chain. A proportionate approval process makes that surface visible.

Information leakage through context, responses and logs

Information can leak before, during or after a call: from the context the model receives to tool responses and logs.

Cybersecurity10 min

MCP security checklist: a central guide for developers and organizations

This guide brings together controls to evaluate MCP servers before and after integrating them. Use it alongside the series’ articles to go deeper into each topic.

AI and agents7 min

Dark factories: what an experiment reveals about agents and architecture

Automating code writing and testing does not, on its own, automate architectural judgment. A case reported by HumanLayer helps show where human review is worth keeping.

AI and agents8 min

A 4-gate workflow: reduce uncertainty before coding with agents

Separating problem, research, architecture and slices turns a broad request into reviewable decisions before an extensive diff is generated.

AI and agents8 min

Context for agents: keeping focus through deliberate compaction

A large context window does not guarantee that an agent uses all the information equally well. Selection and summarization must preserve evidence and useful state.

AI and agents9 min

12 Factor Agents: patterns for agents that reach production

HumanLayer’s 12 factors propose agents built from explicit software, structured tools, controlled state and human pauses when needed.